Meaningful ciphertext encryption for color image via 2D-CICCM and stairwise compression Authors: Shiwei Jing, Jianjun Li Published: 2026 ABSTRACT With the widespread dissemination of digital media, the security of image data has attracted increasing attention. However, existing colour image meaningful ciphertext encryption schemes exhibit low ciphertext quality and poor reconstructed image quality. To address these issues, this paper proposes a colour image encryption scheme for meaningful ciphertext based on a two-dimensional coupled infinite collapsing chaotic mapping (2D-CICCM) and stairwise compression. First, the plaintext image is processed using stairwise compression followed by bit-level permutation to generate a noise-like ciphertext. The generated ciphertext is then embedded into a cover image to produce a meaningful ciphertext image. Unlike conventional compressive sensing methods, which usually neglect inter-channel redundancy in colour images, the proposed stairwise compression jointly exploits both spatial and inter-channel redundancies, thereby significantly improving the reconstruction quality of the decrypted image. Considering the Gaussian-like distribution of measurement values and the characteristics of GLSB-based embedding with correction, a histogram rearrangement and embedding-value adjustment strategy is further designed. This strategy minimises the distortion introduced to the cover image under given embedding constraints, leading to optimal meaningful ciphertext image quality. The pseudo-random sequences required in the encryption process are generated by the 2D-CICCM, which exhibits uniformly distributed attractors, relatively large Lyapunov exponents, and strong randomness, thus satisfying the requirements of cryptographic pseudo-random number generators. Experimental results demonstrate that the proposed scheme achieves favourable performance in terms of ciphertext quality, decryption accuracy, and security, validating its feasibility and effectiveness. The code is available at https://github.com/jsw1995/MCCIE_2D-CICCM. Introduction With the continuous advancement of society and technology, individuals’ demand for collecting, sharing, and storing information has grown dramatically, while the convenience of related operations has also significantly improved. However, this also creates more opportunities for criminals to obtain private data, prompting extensive research on information security by many scholars. Due to their high information density and intuitive nature, images are widely used, and therefore their security is of particular importance. Common image encryption techniques include optical image encryption1,2, quantum image encryption3,4, digital permutation–diffusion encryption5,6, DNA coding7,8, RNA-based operations9,10, bit-plane encryption11,12, among others. These methods convert the plaintext image into a noise-like ciphertext, effectively concealing the visual information. However, the noise characteristics make such ciphertext easier to detect and potentially intercept during transmission. To mitigate this problem, Bao13 introduced meaningful ciphertext encryption, wherein encrypted data are embedded into a cover image, making the ciphertext visually indistinguishable from the cover and thus significantly reducing the likelihood of being detected. Depending on the embedding domain, meaningful ciphertext encryption can be categorized into transform-domain and spatial-domain methods. Transform-domain embedding is commonly implemented in the wavelet transform domain14,15, discrete cosine transform domain16,17, fractional Fourier domain18, skew transform domain19, and U-transform domain20. Among these, integer wavelet transform-based methods are widely adopted due to their reversibility, as no quantization error is introduced during embedding. Typical spatial-domain schemes include least significant bit (LSB) embedding21,22, LSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} correction15,23, and matrix coding24,25. Although spatial-domain methods generally produce ciphertext with higher visual quality than transform-domain methods, the embedding process is independent of the noise-like ciphertext generated through conventional encryption and fails to exploit the known statistical distribution of the ciphertext, limiting further improvements in visual quality26. To address this, Jing27,28 introduced histogram shifting to enhance ciphertext quality, achieving quality far exceeding the theoretical bounds of LSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} correction. Image ciphertexts are typically large in size, and meaningful ciphertext encryption further increases the ciphertext volume relative to noise-like encryption. Compressive sensing (CS) not only reduces plaintext data volume for efficient transmission but also inherently satisfies the confusion and diffusion principles required in cryptography29, leading to its widespread use in image encryption. Conventional CS-based compression typically involves three stages: sparsification, scrambling to uniformly distribute sparse coefficients, and measurement sampling. Common sparsification operators include wavelet transforms30–32, discrete cosine transform20, and singular value decomposition33. CS variants also include block-wise compression34–36, one-dimensional compression30,32, two-dimensional compression37,38, and parallel compression for acceleration25,39. Many optimization strategies have been proposed to further improve reconstruction fidelity, such as adaptive sparse thresholding to select suitable thresholds for different images28, adaptive sampling rate assignment for image blocks40, adaptive sparse basis selection33, and optimized measurement matrices41,42. However, most CS-based methods are designed for grayscale images; color images are typically treated as three independent grayscale channels without exploiting inter-channel correlations, resulting in suboptimal reconstruction quality. Chaotic systems exhibit extreme sensitivity to initial conditions and control parameters, and their generated sequences possess ergodicity and pseudo-randomness. When used to produce random control sequences for image encryption, chaos-based mechanisms not only enhance security but also reduce key size, enabling efficient key distribution while improving key sensitivity43. As a result, a large number of chaotic systems have been adopted for image encryption. One-dimensional chaotic maps such as logistic and sine maps offer structural simplicity and fast computation but their low-dimensional trajectories and limited variables make it easier to deduce initial states44, compromising security. To enhance unpredictability while retaining structural simplicity, some researchers have constructed two-dimensional hyperchaotic systems by cascading or coupling one-dimensional chaotic maps. For example,45 formed a two-dimensional hyperchaos by multiplicatively coupling two infinite-collapse chaotic maps, while28 coupled infinite-collapse and logistic maps under a square-root structure. Multiplicative coupling schemes require carefully designed combinations of low-dimensional maps to ensure value-range constraints. Other works46–48 use additive coupling between low-dimensional maps or between a chaotic map and a simple function, with trigonometric functions applied for range control; however, trigonometric functions often lead to non-uniform value distributions. Similar studies49–51 adopt additive coupling and modular operations, while52 employs logarithmic functions to increase complexity. Additional efforts46,50 introduce non-cyclic infinite operations to further strengthen chaotic behavior. Based on the above discussion and analysis, this study proposes a color image encryption scheme with meaningful ciphertext, which achieves visually secure ciphertext transmission by integrating the 2D-CICCM with a stairwise compression mechanism. The main innovations are as follows:The proposed 2D-CICCM hyperchaotic system is constructed by coupling two infinite-collapse chaotic maps to enhance the dimensionality. While maintaining structural simplicity, it exhibits higher dynamical complexity. Analysis of the Lyapunov exponents, attractors, permutation entropy, and sample entropy demonstrates strong pseudorandomness, confirming that the system satisfies the cryptographic requirements for secure pseudorandom sequence generation in image encryption.To exploit the strong inter-channel correlation of colour images, a stairwise compression strategy is introduced. First, a three-dimensional discrete cosine transform (3D-DCT) is applied to jointly sparsify the spatial and channel dimensions of the RGB image. Then, decreasing sampling rates are assigned according to the energy distribution among channels. This allocation preserves more measurements for high-energy components, thereby ensuring improved reconstruction quality during decryption.To improve ciphertext quality, a histogram rearrangement and embedding-value adjustment strategy is proposed. This strategy fully exploits the fact that compressive sensing measurements approximately follow a Gaussian distribution. Histogram rearrangement maximises the proportion of 0 bits in the binarised measurements, while embedding-value adjustment swaps the decomposed embedding values 2 and 3 (i.e. 2→3 and 3→2). Under the GLSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} corrected embedding conditions, these two operations jointly minimise the modifications introduced to the cover image, thereby achieving optimal visual quality of the ciphertext.The remainder of this paper is organized as follows. Section "Two-dimensional coupled infinite collapsing chaotic mapping" introduces the 2D-CICCM and presents a comparative analysis of its chaotic performance. Section "Three-channel stairwise compressed sensing" describes the stairwise sampling strategy and evaluates its effectiveness. Section "Histogram rearrangement and embedding-value adjustment" details the processes of histogram reorganization and embedded-value adjustment, along with an analysis of their underlying mechanisms. Section "Encryption and decryption process" presents the complete encryption and decryption procedures of the proposed system. Section "Experiment analysis" reports experimental results and validates its security and practicality through comparative analysis. Finally, Sect. "Conclusion" concludes the paper with a comprehensive summary. The proposed 2D-CICCM hyperchaotic system is constructed by coupling two infinite-collapse chaotic maps to enhance the dimensionality. While maintaining structural simplicity, it exhibits higher dynamical complexity. Analysis of the Lyapunov exponents, attractors, permutation entropy, and sample entropy demonstrates strong pseudorandomness, confirming that the system satisfies the cryptographic requirements for secure pseudorandom sequence generation in image encryption. The proposed 2D-CICCM hyperchaotic system is constructed by coupling two infinite-collapse chaotic maps to enhance the dimensionality. While maintaining structural simplicity, it exhibits higher dynamical complexity. Analysis of the Lyapunov exponents, attractors, permutation entropy, and sample entropy demonstrates strong pseudorandomness, confirming that the system satisfies the cryptographic requirements for secure pseudorandom sequence generation in image encryption. To exploit the strong inter-channel correlation of colour images, a stairwise compression strategy is introduced. First, a three-dimensional discrete cosine transform (3D-DCT) is applied to jointly sparsify the spatial and channel dimensions of the RGB image. Then, decreasing sampling rates are assigned according to the energy distribution among channels. This allocation preserves more measurements for high-energy components, thereby ensuring improved reconstruction quality during decryption. To exploit the strong inter-channel correlation of colour images, a stairwise compression strategy is introduced. First, a three-dimensional discrete cosine transform (3D-DCT) is applied to jointly sparsify the spatial and channel dimensions of the RGB image. Then, decreasing sampling rates are assigned according to the energy distribution among channels. This allocation preserves more measurements for high-energy components, thereby ensuring improved reconstruction quality during decryption. To improve ciphertext quality, a histogram rearrangement and embedding-value adjustment strategy is proposed. This strategy fully exploits the fact that compressive sensing measurements approximately follow a Gaussian distribution. Histogram rearrangement maximises the proportion of 0 bits in the binarised measurements, while embedding-value adjustment swaps the decomposed embedding values 2 and 3 (i.e. 2→3 and 3→2). Under the GLSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} corrected embedding conditions, these two operations jointly minimise the modifications introduced to the cover image, thereby achieving optimal visual quality of the ciphertext. To improve ciphertext quality, a histogram rearrangement and embedding-value adjustment strategy is proposed. This strategy fully exploits the fact that compressive sensing measurements approximately follow a Gaussian distribution. Histogram rearrangement maximises the proportion of 0 bits in the binarised measurements, while embedding-value adjustment swaps the decomposed embedding values 2 and 3 (i.e. 2→3 and 3→2). Under the GLSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} corrected embedding conditions, these two operations jointly minimise the modifications introduced to the cover image, thereby achieving optimal visual quality of the ciphertext. Two-dimensional coupled infinite collapsing chaotic mapping Traditional one-dimensional discrete chaotic maps have simple structural forms but their trajectories can be easily estimated and predicted using modern estimation techniques. In contrast, high-dimensional continuous hyperchaotic systems possess stronger chaotic characteristics but incur significantly higher computational cost27. This paper proposes a novel 2D-CICCM, which is constructed based on simple one-dimensional chaos and achieves two-dimensional hyperchaos through cascading coupling. As a result, it retains a simple structure while exhibiting desirable chaotic performance. Mathematical principle Infinite-collapse chaotic maps are algebraically simple, easy to implement, and exhibit strong randomness. Due to these advantages, they have been widely adopted in chaotic secure communication systems. Equation (1) provides the mathematical definition of infinite-collapse chaotic maps.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_{n+1} = \sin (\mu / x_n) \end{aligned}$$\end{document}The chaotic regime of this mapping is given by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu \ne 0,$$\end{document} and the sequence lies in the interval [-1,1]. By coupling two infinite-collapse chaotic maps, a two-dimensional discrete hyperchaotic system is constructed, as shown in Eq. (2), referred to as the two-dimensional coupled infinite-collapse chaotic mapping (2D-CICCM). To address the non-uniform output distribution of the sine function, a modulus operation is applied to the outermost layer to normalize the output to a specific interval. In Eq. (2), π denotes the circular constant, e denotes Euler’s number, M is the modulus factor, and the sequences \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{x_n\}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{y_n\}$$\end{document} take values in the interval [0, M].2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \left\{ \begin{array}{l} x_{n+1}=(a^\pi /(y_{n}\times sin \left( b^e / x_{n}\right) ))\bmod M \\ y_{n+1}=(a^\pi /(x_{n}\times sin \left( b^e / y_{n}\right) ))\bmod M \end{array}\right. \end{aligned}$$\end{document} Performance analysis To verify that the proposed 2D-CICCM exhibits good ergodicity and hyperchaotic behavior, seven recently improved two-dimensional chaotic maps are selected for comparative analysis. Table 1 summarizes the mathematical formulations and corresponding parameter ranges of these improved chaotic systems for reference.Table 1Different two-dimensional chaos.Ref2D chaotic mapsControl parameters2D-SLC24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{l} x_{n+1}=\sin (a\pi x_{n} + b y_{n}\cos (\pi y_{n})) \\ y_{n+1}=\sin (b\pi y_{n} + a x_{n}\cos (\pi x_{n})) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c} a\ne 0 \\ b\ne 0 \end{array}$$\end{document}2D-IICM28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{l} x_{n+1}=\sqrt{1-by_{n}^2} \sin (a/x_{n}) \\ y_{n+1}=\sqrt{1-bx_{n}^2} \sin (a/y_{n}) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c} a\ne 0 \\ b \in [0,1] \end{array}$$\end{document}2D-SCIM53\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{c} x_{n+1}=sin(\pi (2.595abx_{n}(1-x_{n}^2)+\sin ((100-b)/y_{n}))) \\ y_{n+1}=sin(\pi (2.595aby_{n}(1-y_{n}^2)+\sin ((100-b)/x_{n}))) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{l} a \in [1,100]\\ b \in [1,100] \end{array}$$\end{document}2D-SHHM54\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{c} x_{n+1}=k1\sin (1-ax_{n}^2+y_{n}) \\ y_{n+1}=k2\sin (bx_{n}) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{l} a\ne 0 \quad b\ne 0 \\ k1\ne 0 \quad k2\ne 0 \end{array}$$\end{document}2D-DDQCM55\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{l} x_{n+1}=\sin (a-(b-4y_{n}^2)^2) \\ y_{n+1}=\sin (b-(a-4x_{n}^2)^2) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{l} a> 0 \\ b > 0 \end{array}$$\end{document}2D-LLSCM52\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{l} x_{n+1}=\sin (\pi (a\log _{2}{(1+x_n)}(1-\log _{2}{(1+y_n)})/x_n^3 )) \\ y_{n+1}=\sin (\pi (a\log _{2}{(1+y_n)}(1-\log _{2}{(1+x_{n+1})})/y_n^3 )) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c} a > 0 \\ \end{array}$$\end{document}2D-IS23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{c} x_{n+1}=(1-r\times y_{n}^2) \times sin(a/\pi \times x_{n} + b/y_{n}^2) \\ y_{n+1}=(1-k\times y_{n}^2) \times sin(b/\pi \times y_{n} + a/x_{n}^2) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{l} a\ne 0 \quad a\ne 0\\ r \in [0,1] \quad k \in [0,1] \end{array}$$\end{document} Table 1Different two-dimensional chaos.Ref2D chaotic mapsControl parameters2D-SLC24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{l} x_{n+1}=\sin (a\pi x_{n} + b y_{n}\cos (\pi y_{n})) \\ y_{n+1}=\sin (b\pi y_{n} + a x_{n}\cos (\pi x_{n})) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c} a\ne 0 \\ b\ne 0 \end{array}$$\end{document}2D-IICM28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{l} x_{n+1}=\sqrt{1-by_{n}^2} \sin (a/x_{n}) \\ y_{n+1}=\sqrt{1-bx_{n}^2} \sin (a/y_{n}) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c} a\ne 0 \\ b \in [0,1] \end{array}$$\end{document}2D-SCIM53\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{c} x_{n+1}=sin(\pi (2.595abx_{n}(1-x_{n}^2)+\sin ((100-b)/y_{n}))) \\ y_{n+1}=sin(\pi (2.595aby_{n}(1-y_{n}^2)+\sin ((100-b)/x_{n}))) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{l} a \in [1,100]\\ b \in [1,100] \end{array}$$\end{document}2D-SHHM54\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{c} x_{n+1}=k1\sin (1-ax_{n}^2+y_{n}) \\ y_{n+1}=k2\sin (bx_{n}) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{l} a\ne 0 \quad b\ne 0 \\ k1\ne 0 \quad k2\ne 0 \end{array}$$\end{document}2D-DDQCM55\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{l} x_{n+1}=\sin (a-(b-4y_{n}^2)^2) \\ y_{n+1}=\sin (b-(a-4x_{n}^2)^2) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{l} a> 0 \\ b > 0 \end{array}$$\end{document}2D-LLSCM52\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{l} x_{n+1}=\sin (\pi (a\log _{2}{(1+x_n)}(1-\log _{2}{(1+y_n)})/x_n^3 )) \\ y_{n+1}=\sin (\pi (a\log _{2}{(1+y_n)}(1-\log _{2}{(1+x_{n+1})})/y_n^3 )) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c} a > 0 \\ \end{array}$$\end{document}2D-IS23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ \begin{array}{c} x_{n+1}=(1-r\times y_{n}^2) \times sin(a/\pi \times x_{n} + b/y_{n}^2) \\ y_{n+1}=(1-k\times y_{n}^2) \times sin(b/\pi \times y_{n} + a/x_{n}^2) \end{array}\right.$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{l} a\ne 0 \quad a\ne 0\\ r \in [0,1] \quad k \in [0,1] \end{array}$$\end{document} Different two-dimensional chaos. Attractors and histograms To visually illustrate the randomness and ergodicity of different two-dimensional chaotic systems, the attractors and histograms of various chaotic maps are presented. A broader and more uniform distribution of attractors and histograms generally indicates stronger randomness and better ergodicity. In this experiment, the initial state is set to (0.2, 0.3), and each chaotic map is iterated 20,000 times to obtain the corresponding attractor distribution, as shown in Fig. 1. The corresponding histograms of the x-component are shown in Fig. 2.Fig. 1Absorber. 2D-SLC with parameters a = 50, b = 50. 2D-IICM with parameters a = 50, b = 0.99. 2D-SCIM with parameters a = 50, b = 50. 2D-SHHM with parameters a = 50, b = 50, K1=100, K2=100. 2D-DDQCM with parameters a = 50, b = 50. 2D-LLSCM with parameters a = 50. 2D-IS with parameters a = 50, b = 50, r=0.36, k=0.74. 2D-CICCM with parameters a = 50, b = 50, M=1.Fig. 2Histogram. Each chaos parameter is set as in Fig. 1. Absorber. 2D-SLC with parameters a = 50, b = 50. 2D-IICM with parameters a = 50, b = 0.99. 2D-SCIM with parameters a = 50, b = 50. 2D-SHHM with parameters a = 50, b = 50, K1=100, K2=100. 2D-DDQCM with parameters a = 50, b = 50. 2D-LLSCM with parameters a = 50. 2D-IS with parameters a = 50, b = 50, r=0.36, k=0.74. 2D-CICCM with parameters a = 50, b = 50, M=1. Histogram. Each chaos parameter is set as in Fig. 1. From Fig. 1, we observe that all improved two-dimensional chaotic maps are able to cover the defined phase domain; however, the 2D-SLC, 2D-SCIM, 2D-SHHM, 2D-DDQCM, 2D-LLSCM, and 2D-IS exhibit noticeable clustering and regular aggregation patterns. In contrast, the 2D-IICM, and the proposed 2D-CICCM present more uniformly distributed attractors, which is further supported by the histogram visualization in Fig. 16. The unevenness in the former class mainly stems from the use of trigonometric functions as range-constraining mechanisms, which inherently introduces non-uniform density. By comparison, the modular normalization adopted in 2D-CICCM alleviates this limitation and produces a more uniform and unpredictable distribution, demonstrating superior ergodicity and randomness. Lyapunov exponents The Lyapunov exponent (LE) is a key indicator for determining whether a nonlinear system is in a periodic, chaotic, or hyperchaotic state. The number of LEs corresponds to the system dimension. A system with one or more positive LEs is classified as chaotic, while a system with multiple positive LEs is considered hyperchaotic. Furthermore, a larger LE generally implies greater chaotic sensitivity and unpredictability. The heat map of the LEs of the proposed 2D-CICCM under different parameters a and b is presented in Fig. 3a, and the comparative analysis with recently reported 2D chaotic maps under varying parameter a is shown in Fig. 3b and c.Fig. 3LE comparison. 2D-SLC with parameters a∈ (0,100), b = 50. 2D-IICM with parameters a∈ (0,100), b = 0.99. 2D-SCIM with parameters a∈ (1,100), b = 50. 2D-SHHM with parameters a∈ (0,100), b = 50, K1=100, K2=100. 2D-DDQCM with parameters a∈ (0,100), b = 50. 2D-LLSCM with parameters a∈ (0,100). 2D-IS with parameters a∈ (1,100), b = 50, r=0.36, k=0.74. 2D-CICCM with parameters a∈ (0,100), b = 50. LE comparison. 2D-SLC with parameters a∈ (0,100), b = 50. 2D-IICM with parameters a∈ (0,100), b = 0.99. 2D-SCIM with parameters a∈ (1,100), b = 50. 2D-SHHM with parameters a∈ (0,100), b = 50, K1=100, K2=100. 2D-DDQCM with parameters a∈ (0,100), b = 50. 2D-LLSCM with parameters a∈ (0,100). 2D-IS with parameters a∈ (1,100), b = 50, r=0.36, k=0.74. 2D-CICCM with parameters a∈ (0,100), b = 50. From Fig. 3a, the proposed 2D-CICCM exhibits two positive LEs across most of the parameter domain, indicating a broad hyperchaotic parameter space. Furthermore, the magnitude of both positive LEs increases as parameters increase, which suggests that the chaotic intensity can be flexibly enhanced through parameter tuning. As shown in Fig. 3b and c, the proposed 2D-CICCM yields consistently higher positive LEs compared with other improved two-dimensional chaotic maps. This demonstrates its stronger randomness characteristics and greater trajectory sensitivity, making the generated chaotic sequences more difficult to predict and therefore more suitable for secure pseudo-random sequence generation. Complexity analysis Sample entropy (SE) is a nonlinear dynamic metric used to assess the complexity of a time series. A higher sample entropy indicates a greater likelihood of generating new patterns within the sequence, reflecting stronger sequence complexity. Permutation entropy (PE) evaluates the randomness of a dynamic system by analysing the ordinal patterns of the time series and constructing the corresponding probability distribution. PE values closer to 1 imply higher randomness of the underlying system. Figures 4 and 5 present the SE and PE results of the compared chaotic sequences. Figures 4 and 5 show that the mean SE of the proposed 2D-CICCM reaches 2.185, which is substantially higher than those of the other improved chaotic maps. Meanwhile, its PE reaches 0.98, which is very close to the ideal value of 1. These results demonstrate that the 2D-CICCM possesses excellent sequence complexity and randomness, thereby confirming its suitability for secure image encryption applications.Fig. 4SE comparison. Each chaos parameter is set as in Fig. 3.Fig. 5PE comparison. Each chaos parameter is set as in Fig. 3. SE comparison. Each chaos parameter is set as in Fig. 3. PE comparison. Each chaos parameter is set as in Fig. 3. To further verify the randomness of the chaotic sequence generated in this paper, the NIST SP800-22 test suite (comprising 15 subtests, significance level α = 0.01, with a P-value greater than 0.01 considered as passing) published by the National Institute of Standards and Technology (NIST) is employed. A sequence of length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^8$$\end{document} is generated using 2D-CICCM. After binarization with a threshold of 0.5, the sequence is divided into 100 groups, each of length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^6$$\end{document} bits. Based on the confidence interval formula \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{p} \pm 3 \sqrt{\hat{p}(1-\hat{p})/m},$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{p}=0.99$$\end{document} and m = 100, the confidence interval is calculated as [0.9602, 1.0198]. As shown in Fig. 6, all subtests yield P-values greater than 0.01, and the passing rates fluctuate within this confidence interval. The results demonstrate that the binary sequence generated in this paper passes all NIST subtests, exhibits strong randomness, and is suitable for use as a pseudo-random number generator in practical applications.Fig. 6NIST SP800-22 standard test. NIST SP800-22 standard test. Three-channel stairwise compressed sensing In existing visually meaningful ciphertext image encryption schemes, compressive sensing is predominantly applied to grayscale images, while colour images are usually processed by treating the three RGB channels as independent grayscale images. In such approaches, each channel is sparsified using a two-dimensional discrete cosine transform (2D-DCT) independently. This channel-wise strategy ignores the strong inter-channel correlation inherent in colour images, which results in inefficient utilisation of sampling resources and consequently limits reconstruction quality under a fixed sampling rate. To address this limitation, we propose a three-channel stairwise compressed sensing framework that jointly exploits spatial and inter-channel correlations. The proposed framework first performs joint sparsification in a three-dimensional transform domain using a 3D-DCT, followed by stairwise sampling with channel-dependent sampling ratios. By allocating sampling resources according to the information content of each channel, the reconstruction performance can be significantly improved. A schematic illustration of the proposed compression process is shown in Fig. 7.Fig. 7Stairwise compressed sensing. Taking an image of size 512× 512 as an example, the overall sampling rate is set to cr=0.25. Stairwise compressed sensing. Taking an image of size 512× 512 as an example, the overall sampling rate is set to cr=0.25. Let the colour plaintext image be denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$im \in \mathbb {R}^{m \times n \times 3},$$\end{document} where the third dimension corresponds to the RGB colour channels. In conventional compressive sensing schemes, a 2D-DCT is independently applied to each colour channel. Equation (3) presents the 2D-DCT transform for the i-th channel.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} dct_{im_i} = \mathbf {2D\text {-}DCT}(im(:,:,i)) = \sqrt{\frac{2}{m}} \sqrt{\frac{2}{n}} \sum _{x=0}^{m-1}\sum _{y=0}^{n-1} im(x,y,i) \cos \!\left[ \frac{\pi (2x+1)u}{2m}\right] \cos \!\left[ \frac{\pi (2y+1)v}{2n}\right] \end{aligned}$$\end{document}The variables u and v run from 0 to m-1 and from 0 to n − 1, respectively. Although the two-dimensional discrete cosine transform (2D-DCT) can effectively compact spatial energy within each channel, this independent processing strategy ignores the strong spectral correlations among colour channels. Consequently, it fails to fully exploit the inherent redundancy of colour images. Therefore, this paper adopts the three-dimensional discrete cosine transform (3D-DCT) given in Eq. (4) to perform sparsification simultaneously in both the spatial and channel dimensions.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} dct_{im} = \mathbf {3D\text {-}DCT}(im) = \sqrt{\frac{2}{m}} \sqrt{\frac{2}{n}} \sqrt{\frac{2}{3}} \sum _{x=0}^{m-1}\sum _{y=0}^{n-1}\sum _{z=0}^{2} im(x,y,z) \prod _{\ell \in \{u,v,w\}} \cos \!\left[ \frac{\pi (2\xi _\ell +1)\ell }{2L_\ell }\right] \end{aligned}$$\end{document}In Eq. (4), the correspondences are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\xi _u,\xi _v,\xi _w)=(x,y,z)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(L_u,L_v,L_w)=(m,n,3).$$\end{document} By jointly transforming the spatial and channel dimensions, the 3D-DCT compacts a large proportion of the signal energy into a small number of low-frequency coefficients along the channel dimension. As a result, the effective sparsity of the transformed representation is significantly reduced compared with independent 2D-DCT-based channel-wise sparsification. According to compressive sensing theory, a signal with sparsity level K can be reliably reconstructed from M measurements if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M \ge C K \log \!\left( \frac{N}{K}\right),$$\end{document} where C is a positive constant and N denotes the signal dimension56. Under a fixed overall sampling rate cr, reducing the effective sparsity directly improves reconstruction performance, which provides the theoretical motivation for adopting 3D-DCT-based sparsification. After 3D-DCT sparsification, the energy distribution across channels becomes highly uneven. This motivates the proposed stairwise sampling strategy, which allocates sampling ratios \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(p_1,p_2,p_3)$$\end{document} satisfying \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_1+p_2+p_3=1,$$\end{document} and assigns more measurements to channels with higher energy concentration. The detailed sampling and measurement construction procedures are summarised in Algorithm 1. Algorithm 1Stairwise compressed sensing Stairwise compressed sensing Three-channel sampling rate The three color channels are sampled using different proportions. Since the subsequent encryption stage requires a total sampling rate of 0.25, we investigate only the allocation schemes satisfying this constraint. Figure 8 illustrates the reconstructed image quality under various distribution strategies for low-frequency and high-frequency channels on the CVG dataset1. As shown in Fig. 8, increasing the sampling ratio of the low- and mid-frequency channels initially improves reconstruction quality, with the optimal performance achieved at the allocation scheme (0.64,0.24). Therefore, the sampling ratios in this work are set to (0.64,0.24,0.12) for the low-, mid-, and high-frequency channels, respectively.Fig. 8Reconstruction of image quality. Horizontal coordinate p1 is the sampling rate ratio of low frequency channels, vertical coordinate p2 is the sampling rate ratio of middle frequency channels, and the sampling ratio of remaining high frequency channels p3=1-p1-p2. Calculate the value only if P1>P2>p3, and substitute the minimum value for the rest of the time. Reconstruction of image quality. Horizontal coordinate p1 is the sampling rate ratio of low frequency channels, vertical coordinate p2 is the sampling rate ratio of middle frequency channels, and the sampling ratio of remaining high frequency channels p3=1-p1-p2. Calculate the value only if P1>P2>p3, and substitute the minimum value for the rest of the time. Ablation study In this section, the effectiveness of the proposed 3D-DCT sparsification and stairwise sampling strategy is mainly evaluated. Two sampling schemes are selected as baselines. The first scheme consists of three steps: sparsification, permutation, and sampling28. The permutation step is designed to uniformly distribute sparse energy over the entire sparse matrix, thereby avoiding excessive concentration of low-frequency components in certain rows. Under the same sampling rate, this strategy reduces information loss and improves the quality of the reconstructed image. The second scheme follows a two-step process, including sparsification and sampling27. In this case, the sampling matrix assigns larger weights to the upper-left region of the sparse coefficient matrix, which preserves more low-frequency information. This design enables fast reconstruction while maintaining high reconstruction quality. To verify the effectiveness of the 3D sparsification and the stairwise sampling strategy, three experiments are conducted. In the first experiment, 3D sparsification is introduced into the first sampling scheme. In the second experiment, YCrCb colour space transformation and stairwise sampling are incorporated into the second sampling scheme. This experiment mainly serves as a supplementary validation of the effectiveness of 3D sparsification. Specifically, the luminance channel Y is treated as the primary channel with a higher sampling rate, while the chrominance channels Cr and Cb are treated as secondary channels with relatively lower sampling rates. The sampling ratios are optimally selected according to Sect. "Three-channel sampling rate", resulting in a ratio of (0.6533, 0.1733, 0.1733). This allocation principle is consistent with that of JPEG compression, as the human visual system is highly sensitive to variations in luminance Y but less sensitive to subtle changes in chrominance Cb and Cr, thereby motivating the use of stairwise sampling. The third experiment corresponds to the proposed method, in which both 3D sparsification and the stairwise sampling strategy are applied to the second sampling scheme. The reconstruction quality of the two baseline methods and their corresponding three enhanced variants on the CVG dataset is reported in Table 2. Figure 9 illustrates the compressive reconstruction results of the portofino image from the dataset.Table 2Reconstruction quality and sampling reconstruction time.RefRef28Ref28 + 3D-DCTRef27Ref27 + YCrCbOurPSNR(dB)/without quantizatio27.956730.640430.214732.168432.4045PSNR(dB)/perform quantization26.519328.478429.691929.899931.5737Time(s)22.874621.78321.25860.80091.0164 Table 2Reconstruction quality and sampling reconstruction time.RefRef28Ref28 + 3D-DCTRef27Ref27 + YCrCbOurPSNR(dB)/without quantizatio27.956730.640430.214732.168432.4045PSNR(dB)/perform quantization26.519328.478429.691929.899931.5737Time(s)22.874621.78321.25860.80091.0164 Reconstruction quality and sampling reconstruction time. Fig. 9Reconstructed images. The first and second rows show the reconstructed images without measurement quantisation and the corresponding magnified views of the red-boxed regions, respectively. The third and fourth rows present the reconstructed images with measurement quantisation and the corresponding magnified views of the red-boxed regions. Reconstructed images. The first and second rows show the reconstructed images without measurement quantisation and the corresponding magnified views of the red-boxed regions, respectively. The third and fourth rows present the reconstructed images with measurement quantisation and the corresponding magnified views of the red-boxed regions. As shown in Table 2, the reconstruction quality in the second column is approximately 2 dB higher than that in the first column, indicating that 3D sparsification provides a significant improvement for the first sampling scheme. By comparing the last two columns with the third column, it can be observed that the proposed stairwise sampling strategy further improves reconstruction performance under both YCrCb colour transformation and 3D sparsification. This improvement demonstrates the effectiveness of the stairwise sampling strategy. The main reason is that the proposed allocation assigns a higher sampling ratio to the luminance channel Y, which is more sensitive to human visual perception, as well as to the lowest-frequency channel after 3D sparsification. Consequently, more informative coefficients are preserved while redundant information is suppressed, leading to improved reconstruction quality. Furthermore, comparing the last two columns shows that the proposed 3D sparsification consistently achieves better reconstruction quality than the YCrCb channel decomposition, regardless of whether measurement quantisation is applied. This observation indicates that joint sparsification in the 3D transform domain is more effective in capturing inter-channel redundancy than colour-space separation. From Fig. 9, it can be seen that the reconstructed images obtained using the proposed sampling method preserve image details most effectively and exhibit the smallest visual difference from the original image. This observation confirms that the proposed 3D-DCT sparsification combined with stairwise sampling not only compresses spatial redundancy but also effectively exploits cross-channel redundancy. As a result, more useful information is retained under the same sampling rate, leading to a significant improvement in reconstruction quality. Overall, the proposed sampling framework achieves superior reconstruction performance while maintaining relatively high compression–reconstruction efficiency, making it well suited for colour meaningful ciphertext image encryption applications. Histogram rearrangement and embedding-value adjustment After stepped compressed sensing, the color plaintext is converted into measurement values that exhibit an approximately Gaussian-like distribution. Based on this observation, this work adopts a histogram-based enhancement mechanism inspired by27,28 to improve the visual quality of the meaningful ciphertext. Unlike27,28, which applies global histogram shifting, this work introduces a histogram reorganization strategy to reduce the modification intensity applied to the cover during the embedding stage. To further improve the ciphertext quality, an Embedded Value Adjustment mechanism is incorporated. Figure 10 illustrates the sequential order of histogram rearrangement, embedding value adjustment, M-ary decomposition, GLSB, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} correction throughout the encryption process. To facilitate the subsequent explanation, the measurement values obtained from compressive sensing are denoted as NC of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m \times n \times 3,$$\end{document} and the original image is denoted as CO of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M \times N \times 3.$$\end{document} Here, we first briefly introduce the GLSB and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} correction proposed in27 to support the subsequent explanation. On this basis, the histogram rearrangement and embedding value adjustment methods are described in detail, and their effectiveness is verified.Fig. 10Embedding process of literature. Embedding process of literature. Preliminaries This paper adopts the method from27 for embedding. The noise-like ciphertext is first decomposed using M-ary decomposition, followed by GLSB and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} correction embedding to obtain the meaningful ciphertext. The following subsections respectively introduce M-ary decomposition, GLSB, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} correction. Binary decomposition is a special case of M-ary decomposition. M-ary decomposition allows the noise-like signal to be decomposed using an arbitrary decomposition base, offering greater flexibility. Since this paper decomposes the noise-like ciphertext into four equal blocks, each containing 2 bits of information, the corresponding M-ary decomposition is expressed using Eq. (5).5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \left\{ \begin{aligned} CA&= \lfloor NC1/64 \rfloor \\ CH&= \lfloor \textrm{mod}(NC1,64)/16 \rfloor \\ CV&= \lfloor \textrm{mod}(NC1,16)/4 \rfloor \\ CD&= \textrm{mod}(NC1,4) \end{aligned} \right. \end{aligned}$$\end{document}The corresponding merging process is shown in Eq. (6).6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} NC1 = CA\times 64 + CH\times 16 + CV\times 4 + CD \end{aligned}$$\end{document}The generalized least significant bit (GLSB) embedding proposed in27 is mainly characterized by its ability to determine the amount of modification to the cover image based solely on the embedded value, without requiring knowledge of the cover image. This allows the ciphertext quality to be improved by adjusting the embedded information. This is the key reason why27 can enhance ciphertext quality through histogram shifting and, in this paper, through histogram rearrangement. The GLSB embedding and extraction processes are expressed using Eqs. (7) and (8), respectively.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} C2(1\!:\!2m,1\!:\!2n,:) = \textrm{mod}(\textrm{mod}(CO(1\!:\!2m,1\!:\!2n,:),4)+EI,\,4)+\lfloor CO(1\!:\!2m,1\!:\!2n,:)/4 \rfloor \times 4 \end{aligned}$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} EI = \textrm{mod}(\textrm{mod}(C2(1\!:\!2m,1\!:\!2n,:),4)-\textrm{mod}(CO(1\!:\!2m,1\!:\!2n,:),4),\,4) \end{aligned}$$\end{document}The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^K$$\end{document} correction given in Eq. (9) transforms the modification values of 3 or -3 in C2 into -1 and 1, respectively, without affecting the extraction process. It is a widely used correction method at the present stage.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} C3(i,j)= {\left\{ \begin{array}{ll} C2(i,j)-4 & C2(i,j)-CO(i,j)>3\ \text {and}\ C2(i,j)-4 \ge 3\\ C2(i,j)+4 & C2(i,j)-CO(i,j)<-3\ \text {and}\ C2(i,j)+4 \le 3\\ C2(i,j) & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document} Histogram rearrangement and embedding value adjustment procedures Table 3 presents the relationship between the embedded information and the amount of modification to the cover image when using the GLSB + \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^k$$\end{document} correction method. It is assumed here that both the embedded information and the noise-like ciphertext are uniformly distributed. The first row represents the values of the cover content after processing by \bmod (CO, 4), and the first column represents the values to be embedded. Each entry consists of two quantities: the resulting pixel change and its probability.Table 3Amount of pixel modification.Change/P012300 1/40 1/40 1/40 1/411 1/41 1/41 1/41 1/422 1/42 1/42 1/42 1/43−1 1/4−1 1/4−1 1/4−1 1/4 Table 3Amount of pixel modification.Change/P012300 1/40 1/40 1/40 1/411 1/41 1/41 1/41 1/422 1/42 1/42 1/42 1/43−1 1/4−1 1/4−1 1/4−1 1/4 Amount of pixel modification. From Table 3, it can be seen that increasing the proportion of embedded values of 0 significantly improves the ciphertext quality. To this end, this paper proposes a histogram rearrangement method. Its main principle is to reassign values to the noise-like ciphertext, which follows a Gaussian-like distribution, so that after binary decomposition, more 0-bits are obtained. Consequently, the probability of an embedded value of 0 increases substantially, and the ciphertext quality is greatly enhanced. Histogram rearrangement first requires determining the sorting order of the 256 pixel values in [0, 255] based on the proportion of 0-bits in their binary representations, which provides guidance for subsequent pixel value reassignment. The order of pixel values from the highest to the lowest proportion of 0-bits can be obtained using Eq. (10) below. This order remains invariant and can be precomputed as a known sequence directly embedded into the program, eliminating the need for repeated calculations and thus saving encryption time. In Eq. (10), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$sum(a_{bin},1)$$\end{document} performs summation along the row dimension, and argsort(a) returns indices that sort a in ascending order.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \left\{ \begin{matrix} a=[0,1,2,\dots ,255]\\ a_{bin}=dec2bin(a)\\ a_{num}=sum(a_{bin},1)\\ index=argsort(a_{num})-1 \end{matrix}\right. \end{aligned}$$\end{document}For the noise-like ciphertext with a Gaussian-like distribution, histogram reconstruction mainly consists of two steps. The first step is to find the highest point of the histogram and shift it to 128 by means of global shifting, as shown in Eq. (11), so as to facilitate subsequent pixel value reassignment. In Eq. (11), Hist(a) finds the histogram distribution of a, and argmax(a) finds the index at the maximum of a.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \left\{ \begin{matrix} h_{NC}=hist(NC)\\ T=127-argmax(h_{NC})\\ NC1=\mod ((NC+T),256) \end{matrix}\right. \end{aligned}$$\end{document}The second step is to reassign values to the noise-like ciphertext, which can be expressed using Eq. (12). In this equation, the operation NC2(NC1==a)=b replaces all positions in NC2 where NC1=a with the value b.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \left\{ \begin{matrix} NC2(NC1==128-i-1)=index(2i)\\ NC2(NC1==128+i)=index(2i+1) & i=0,1,2,\dots ,127 \end{matrix}\right. \end{aligned}$$\end{document}After histogram rearrangement, the proportion of 0-bits increases. Following M-ary decomposition, the proportions of 0, 1, 2, and 3 in descending order are 0, 1, 2, and 3, respectively. Although the increased proportions of 0 and 1 can effectively improve ciphertext quality, as shown in Table 3, the modification amount for a decomposed value of 2 is larger than that for a value of 3. Therefore, it is necessary to further increase the occurrence probability of 3 while reducing that of 2. An intuitive and simple approach is to swap the decomposed values 2 and 3, which constitutes the embedding value adjustment. The process can be expressed using Eq. (13), where the forward and inverse operations are identical.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \left\{ \begin{matrix} EI = EI1 \\ EI(EI1 == 2) = 3 \\ EI(EI1 == 3) = 2 \end{matrix} \right. \end{aligned}$$\end{document} Effectiveness analysis of histogram rearrangement and embedding value adjustment As shown in Table 3, under the condition of a uniformly distributed noise-like signal, the theoretical PSNR of the ciphertext is given by Eq. (14) and is approximately 46.3699 dB.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} PSNR = 10 \times \log _{10} \left( \frac{255^{2}}{ \frac{1}{4} \times 0^{2} + \frac{2}{4} \times 1^{2} + \frac{1}{4} \times 2^{2} } \right) \approx 46.3699\,\text {dB} \end{aligned}$$\end{document}This value also represents the theoretical upper bound attainable by existing LSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^k$$\end{document} embedding schemes15,23. For noise-like ciphertexts following a Gaussian-like distribution, as in Fig. 11a, related works27,28 employ histogram shifting to increase the proportion of zero modifications, thereby improving visual quality. However, histogram shifting is a global, single-direction shift (Fig. 11b), which cannot achieve the optimal histogram distribution. To address this, we introduce histogram reorganization, which explicitly reshapes the distribution into a more optimal structure. In Fig. 11a–c, the proportions of binary 1-bits are 0.5280, 0.3676, and 0.3076, respectively. Equations (15), (16), and (17) provide the theoretical PSNR values for the corresponding cases, which reach 46.2985 dB, 46.9957 dB, and 47.4880 dB, respectively. The results indicate that histogram reconstruction achieves a more significant improvement than histogram shifting.Fig. 11Histogram distribution.15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} PSNR_{a}&= 10 \times \log _{10} \left( \frac{255^{2}}{ 0.4720^{2} \times 0^{2} + (0.4720 \times 0.5280 + 0.5280^{2}) \times 1^{2} + 0.5280 \times 0.4720 \times 2^{2} } \right) \approx 46.2985\,\text {dB} \end{aligned}$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} PSNR_{b}&= 10 \times \log _{10} \left( \frac{255^{2}}{ 0.6318^{2} \times 0^{2} + (0.6318 \times 0.3682 + 0.3682^{2}) \times 1^{2} + 0.3682 \times 0.6318 \times 2^{2} } \right) \approx 46.9957\,\text {dB} \end{aligned}$$\end{document}17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} PSNR_{c}&= 10 \times \log _{10} \left( \frac{255^{2}}{ 0.6924^{2} \times 0^{2} + (0.3076 \times 0.6924 + 0.3076^{2}) \times 1^{2} + 0.6924 \times 0.3076 \times 2^{2} } \right) \approx 47.4880\,\text {dB} \end{aligned}$$\end{document}In the non-uniform distribution shown in Fig. 11, the proportion of 0-bits exceeds that of 1-bits. Consequently, the probability of obtaining a value of 2 is significantly higher than that of obtaining a value of 3. By swapping the values 2 and 3, the likelihood of generating a change of -1 increases, thereby further enhancing the visual quality. Therefore, this paper introduces an embedding value adjustment method. Under the distribution shown in Fig. 11c, the theoretical PSNR after embedding value adjustment is calculated using Eq. (18), reaching 49.0759 dB. Compared with 47.4880 dB before adjustment, this indicates that the proposed embedding value adjustment effectively improves ciphertext quality.18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} PSNR_{d}&= 10 \times \log _{10} \left( \frac{255^{2}}{ 0.7261^{2} \times 0^{2} + (2 \times 0.3076 \times 0.6924) \times 1^{2} + 0.3076^{2} \times 2^{2} } \right) \approx 49.0759\,\text {dB} \end{aligned}$$\end{document}In summary, the proposed histogram reorganization and embedding value adjustment strategies provide an essential enhancement to ciphertext quality beyond conventional histogram shifting and LSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^k$$\end{document} embedding. Histogram shifting27,28 performs a global shift operation that primarily increases the proportion of zero modifications, thereby improving PSNR to some extent. However, its global nature inherently prevents it from reaching the optimal distribution of noise-like ciphertext values. In contrast, the proposed histogram reorganization redistributes the cover values to approximate a distribution that minimizes the expected distortion during GLSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^k$$\end{document} correction. Furthermore, the proposed embedding value adjustment addresses another limitation of conventional LSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^k$$\end{document} embedding, where the correction operation maps a cover difference of 3 to −1. While this mapping minimizes distortion under uniform binary distribution assumptions, it becomes suboptimal when the 0-bit proportion significantly exceeds that of 1-bit, as the occurrence probability of a correction producing a value of 2 becomes much higher than that of 3. By interchanging the correction mapping and allowing the more probable event to correspond to the smaller modification magnitude, the expected distortion is reduced, and a higher theoretical PSNR can be achieved. Combined with histogram reorganization, this yields the best theoretical performance attainable under LSB+\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^k$$\end{document} embedding, confirming the superiority of the proposed method for noise-like ciphertext generation. Histogram distribution. Encryption and decryption process Let the plaintext image be denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P \in \mathbb {R}^{m \times n \times 3},$$\end{document} and the cover image be denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CO \in \mathbb {R}^{M \times N \times 3}.$$\end{document} The secret key is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H1 = \{h1_{1}, h1_{2}, \dots , h1_{256}\}.$$\end{document} The global sampling rate is set to 0.25, and the sampling allocation ratios for the three color channels (R, G, and B) are configured as 0.64, 0.24, and 0.12, respectively. The overall encryption and decryption framework is illustrated in Fig. 12. During the encryption stage, the sender first performs compression and encryption operations on the plaintext image to generate a noise-like ciphertext along with a dynamic key. The ciphertext and the associated key information are then transmitted to the receiver through a communication channel. At the receiver side, the transmitted ciphertext is processed through the inverse decryption and reconstruction procedures to recover the original plaintext information. To enhance security and make the distribution of embedded values more uniform across the cover image, bit scrambling and cover scrambling are introduced into the process described in Sect. "Histogram rearrangement and embedding-value adjustment". The detailed encryption and decryption procedures are explained step by step below.Fig. 12Encryption and decryption process. The black box represents the encryption process, the green box represents the decryption process, and the yellow box indicates the shared operations between encryption and decryption. Encryption and decryption process. The black box represents the encryption process, the green box represents the decryption process, and the yellow box indicates the shared operations between encryption and decryption. Encryption process As illustrated in Fig. 12, the proposed encryption process consists of four main stages: chaotic pseudo-random sequence generation, stairwise compressive sensing, noise-like ciphertext encryption via histogram reorganization and bit scrambling, and data embedding. To enhance plaintext sensitivity and strengthen resistance against chosen-plaintext and chosen-ciphertext attacks, the proposed scheme incorporates both temporal information and plaintext hash values into the dynamic key generation mechanism. The detailed encryption procedure is described as follows: Dynamic key generation and chaotic sequence generation Step 1: The dynamic key generation follows Algorithm 2. Algorithm 2Dynamic key generation Dynamic key generation Step 2: Parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(a_1,b_1,x_1,y_1)$$\end{document} are substituted into (2) and iterated 3mn+1000 times; the first 1000 iterations are discarded to eliminate transient effects, resulting in sequences r1, r2 for stairwise sampling. Similarly, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(a_2,b_2,x_2,y_2)$$\end{document} yield (r3, r4) of length \max (4m,6n) for bit scrambling, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(a_3,b_3,x_3,y_3)$$\end{document} yield (r5, r6) of length \max (3M,N) for ciphertext embedding. Stairwise compressed sensing The plaintext P and sequences (r1, r2) are fed into Algorithm 1 to obtain the measurement matrix M. Noise-like ciphertext ecryption Step 1: Extract the histogram control sequence index and apply it together with the measurement matrix M to Eqs. (11) and (12) to perform histogram rearrangement, obtaining the rearranged matrix M2. Step 2: The sequence (r3, r4) is converted into cyclic shift parameters using Eqs. (19) and (20).19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} R3 = \mod (r3 \times 10^4,\, 3n) \end{aligned}$$\end{document}20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} R4 = \mod (r4 \times 10^4,\, 2m) \end{aligned}$$\end{document} Step 3: Convert the histogram-shifted matrix M2 into its binary representation. To balance the number of rows and columns, a reshaping operation is performed.21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} bin\_M = reshape(dec2bin(M2), [4m,\,2k n]) \end{aligned}$$\end{document}The operation reshape(a, [m, n]) reformats the matrix a into a new matrix of dimensions m × n. Step 4: Bit scrambling is applied to the binary matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$bin_M$$\end{document} by performing cyclic shifts along both the row and column dimensions.22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} bin\_M(i,:)=circshift(bin\_M(i,:),r1(i)),\quad i=1,2,\dots ,4m \end{aligned}$$\end{document}23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} bin\_M(:,i)=circshift(bin\_M(:,i),r2(i)),\quad i=1,2,\dots ,2k n \end{aligned}$$\end{document}circshift(a, b) performs a circular right shift of vector a by b positions. Step 5: Binary-to-decimal conversion to obtain noise-like ciphertext NC.24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} NC = reshape(bin2dec(reshape(bin\_M,[mnk, 8])),[m,n,k]) \end{aligned}$$\end{document} Meaningful ciphertext generation Step 1: M-ary decomposition is applied to the noise-like ciphertext according to Eq. (5) to obtain four components, which are then concatenated using Eq. (25) to form the matrix EI1.25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} EI1 =\begin{bmatrix} CA& CH \\ CV& CD \end{bmatrix} \end{aligned}$$\end{document} Step 2: Embedding value adjustment is applied to the concatenated matrix EI1 according to Eq. (13) to obtain the adjusted matrix E1. Step 3: The sequences (r5, r6) are converted into cyclic shift control sequences according to Eqs. (26) and (27), respectively. After concatenating the channels of CO along rows, row and column scrambling is performed using Eqs. (22) and (23). Finally, the scrambled matrix CO2 is obtained by concatenating the three RGB channels.26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} R5 = \mod (r5 \times 10^4,\, N) \end{aligned}$$\end{document}27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} R6 = \mod (r6 \times 10^4,\, 3M) \end{aligned}$$\end{document} Step 4: The GLSB and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^k$$\end{document} correction embedding operations are applied to CO2 and the value to be embedded E1 according to Eqs. (7) and (9) to obtain C1. Step 5: Negative signs are added to R5 and R6, and inverse scrambling is applied to C1 using Eqs. (22) and (23) to obtain the meaningful ciphertext C. Decryption process The decoder requires not only the ciphertext but also the corresponding dynamic key. The key can be securely transmitted through public networks using traditional asymmetric cryptographic schemes such as RSA, or through dedicated private channels. The decryption procedure consists of four stages, namely sequence generation, information extraction, noise-like ciphertext decryption, and reconstruction. Chaotic sequence generation The decryption party substitutes the obtained dynamic key into Step 2 of the third subsection of the encryption process to obtain the chaotic sequences r1, r2, r3, r4, r5, r6. Information extraction Step 1: Apply Steps 3 of the meaningful ciphertext generation process to the ciphertext C and the cover CO to obtain C2 and CO2. Step 2: The embedded information EI1 is extracted by substituting C2 and CO2 into Eq. (8). Step 3: The inverse embedding value operation is applied to EI1 according to Eq. (13) to obtain EI. Step 4: EI is decomposed into four parts according to Eq. (28), and then merged using Eq. (6) to obtain the noise-like ciphertext NC.28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \left\{ \begin{aligned} CA&= EI(1\!:\!m/2,1\!:\!n/2)\\ CH&= EI(1\!:\!m/2,n/2+1\!:\!n)\\ CV&= EI(m/2+1\!:\!m,1\!:\!n)\\ CD&= EI(m/2+1\!:\!m,n/2+1\!:\!n) \end{aligned} \right. \end{aligned}$$\end{document} Noise-like ciphertext decryption Step 1: The inverse bit scrambling is performed by applying negative shift parameters r1 and r2 to undo the Noise-like ciphertext ecryption in Steps 2-5, recovering M2. Step 2: Retrieve the histogram distribution index index and perform inverse histogram reconstruction.29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \left\{ \begin{matrix} M1(M2==index(2i))=128-i-1\\ M1(M2==index(2i+1))=128+i & i=0,1,2,\dots ,127 \end{matrix}\right. \end{aligned}$$\end{document}Step 3: Apply the inverse shift to recover the measurement value M:30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} M =\mod (M1-T, 256) \end{aligned}$$\end{document} Reconstruction Step 1: During reconstruction, each channel is reconstructed separately according to the sampling rate allocated to it to obtain the 3D-DCT sparse matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{dct}$$\end{document}. The reconstruction algorithm employs the Nested Smoothing L0 Norm (NSL0) algorithm, with parameters set as follows: the minimum Gaussian kernel width \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _{\min }$$\end{document} is 0.01, the sigma decay factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _{\text {decrease-factor}}$$\end{document} is 0.05, and the convergence threshold ξ is 0.01. Step 2: The inverse three-dimensional discrete cosine transform (3D-IDCT) is applied to the sparse matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$im_{dct}$$\end{document} to obtain the final decrypted image RP.31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} RP = \mathbf {3D\text {-}IDCT}(RP) \end{aligned}$$\end{document} Experiment analysis In the experiments, the CVG dataset is adopted as the test benchmark. The security key is configured as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$sec_{key}$$\end{document}=8d5ab8ba5340 fce4420829ad5d12a0e45dacb0858544163d04c1d02b73e3697d. The overall sampling rate is set to 0.25, and the channel-wise sampling allocation ratios are assigned as 0.64, 0.24, and 0.12 for the three color channels, respectively. Several representative methods are chosen for evaluation, including schemes with high ciphertext visual quality15,27,28 and those with superior reconstruction performance24,57. Among them, Ref.27 adopts two different embedding strategies with comparable ciphertext quality; therefore, the spatial-domain embedding scheme is selected for comparison. For Ref.57, the number of inner-loop iterations is set to four. Encryption and decryption effect Several images from the dataset are selected as plaintexts and cover images, with their encryption and decryption results presented in Fig. 13. In each example, the meaningful ciphertext and the decrypted image additionally include a difference map displayed at their bottom-right corners. To facilitate visual inspection of the ciphertext distortion and reconstruction accuracy, the difference maps are magnified by 30.Fig. 13Encryption and decryption result. For the visually meaningful ciphertext and the decrypted image, the bottom-right corners present the difference images with respect to the cover image and the plaintext, respectively, where the differences are amplified by a factor of 30. Encryption and decryption result. For the visually meaningful ciphertext and the decrypted image, the bottom-right corners present the difference images with respect to the cover image and the plaintext, respectively, where the differences are amplified by a factor of 30. As observed in Fig. 13, the noise-like ciphertext exhibits noise-like characteristics, lacking both structurally identifiable patterns and semantically interpretable content. This effectively prevents adversaries from inferring the original information from the ciphertext alone at the information level, thereby providing fundamental security assurance for data transmission. In contrast, the visually meaningful ciphertext is nearly indistinguishable from the carrier image, with the corresponding difference image appearing almost black, indicating that the discrepancy between the two is confined to an extremely small pixel amplitude range and is therefore difficult to discern visually. Such high-fidelity concealment demonstrates that the proposed method can successfully disguise easily detectable noise-like ciphertext as natural images, significantly reducing the probability of detection or interception during transmission and enhancing the transmission security of covert channels. Meanwhile, the decrypted images clearly recover the plaintext information with well-preserved details and no evident distortion or artifacts, indicating that the decryption process achieves high reconstruction quality in restoring the original content. Overall, the proposed encryption algorithm ensures both information security and complete information recovery. Conventional security analysis Here, the Toucan and Portofino images from the second row of Fig. 13 are selected as the plaintext and cover image, respectively, for the following experiments. Key sensitivity analysis The proposed scheme contains twelve dynamic keys, denoted as Dynkey = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${a_1,b_1,x_1,y_1; a_2,b_2,x_2,y_2; a_3,b_3,x_3,y_3},$$\end{document} which are used to generate the measurement matrices and control sequences for compression, bit scrambling, and embedding scrambling. To verify the sensitivity of the keys, slight perturbations were introduced to the dynamic keys during decryption, and the resulting images are shown in Fig. 14.Fig. 14Decrypted images under slightly modified keys. Decrypted images under slightly modified keys. As shown in Fig. 14, regardless of whether the key is used to generate the measurement matrix for controlling the compressive sampling process or to generate the scrambling control sequence for altering the pixel arrangement order, even an extremely slight modification to the key causes the decrypted image to completely degenerate into noise-like invalid information, from which no intelligible content can be discerned. This phenomenon fully demonstrates that the proposed algorithm exhibits strong sensitivity across all key application stages—a minor deviation in the key leads to complete failure of the decryption process. Such high key sensitivity ensures that the algorithm can effectively resist attacks such as brute-force cracking and key guessing. Even if an attacker obtains parameters close to the correct key, they cannot recover the plaintext information, thereby providing robust security assurance for the system. Key space analysis It is generally accepted that a key space larger than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^{100}$$\end{document} is sufficient to resist brute-force attacks58, a criterion that provides a basic quantitative basis for the security of encryption schemes. From the above key sensitivity analysis, it can be observed that even when the precision of the twelve dynamic keys is reduced to within \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-14}$$\end{document}, each key remains highly sensitive, with a minor numerical deviation leading to decryption failure. Based on this characteristic, the key space of the proposed scheme can reach \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{14\times 12} \approx 2^{558}$$\end{document}, far exceeding the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^{100}$$\end{document} security threshold. Such a vast key space implies that even with substantial computational resources, an attacker cannot feasibly crack the key through brute-force enumeration within a limited timeframe. Therefore, the proposed algorithm provides sufficient security redundancy at the key management level, effectively resisting brute-force attacks and meeting the practical requirements of high-security application scenarios.It is generally accepted that a key space larger than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^{100}$$\end{document} is sufficient to resist brute-force attacks58, a criterion that provides a basic quantitative basis for the security of encryption schemes. From the above key sensitivity analysis, it can be observed that even when the precision of the twelve dynamic keys is reduced to within \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-14}$$\end{document}, each key remains highly sensitive, with a minor numerical deviation leading to decryption failure. Based on this characteristic, the key space of the proposed scheme can reach \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{14\times 12} \approx 2^{558}$$\end{document}, far exceeding the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2^{100}$$\end{document} security threshold. Such a vast key space implies that even with substantial computational resources, an attacker cannot feasibly crack the key through brute-force enumeration within a limited timeframe. Therefore, the proposed algorithm provides sufficient security redundancy at the key management level, effectively resisting brute-force attacks and meeting the practical requirements of high-security application scenarios. Robustness analysis During transmission, ciphertext is inevitably subject to channel noise interference or packet loss, resulting in incomplete or distorted information at the receiver end. Therefore, robustness becomes an important criterion for evaluating the practical applicability of encryption schemes. To comprehensively assess the performance of the proposed method under complex transmission conditions, this paper introduces salt-and-pepper noise of different densities and Gaussian noise with zero mean and different variances to simulate pixel-level perturbations, and employs cropping of different block sizes to simulate data loss. After applying such disturbances to the ciphertext, the recovery quality of the decrypted image is observed. The corresponding decryption results are shown in Fig. 15.Fig. 15Decryption results under clipping and noise attacks. Decryption results under clipping and noise attacks. As shown in Fig. 15, even when the ciphertext suffers from partial content loss during transmission or storage, or is disturbed by varying levels of noise, the decrypted image remains clearly recognizable in terms of its main content, without overall information failure caused by localized data corruption. This characteristic demonstrates that the proposed scheme adequately accounts for non-ideal factors that may be encountered in real-world channels, exhibiting strong fault tolerance. Specifically, whether facing spatial information loss due to cropping attacks or pixel value perturbations caused by superimposed noise, the decryption process effectively suppresses the propagation of distortion and ensures the faithful recovery of essential semantic information. These results confirm that the method possesses strong robustness in complex environments, satisfying the practical reliability requirements of covert communication. Statistical characteristics analysis Histogram Histogram analysis can clearly and intuitively reflect the changes in pixel value distribution from plaintext to ciphertext. For noise-like encryption, the cipher-image histogram should differ significantly from that of the plaintext in order to conceal information effectively, whereas meaningful ciphertext should maintain a histogram similar to the cover image for visual plausibility. The histograms of the encrypted and decrypted results are presented in Fig. 16.Fig. 16Histogram. Histogram. As shown in Fig. 16, the histogram of the noise-like ciphertext differs significantly from that of the plaintext, exhibiting a more uniform pixel value distribution with the original statistical features completely disrupted. This indicates that the compression and bit-scrambling operations fundamentally alter the statistical distribution of plaintext pixel values, eliminating correlations and structural regularities between adjacent pixels. Consequently, the plaintext information is effectively concealed, making it difficult for attackers to infer the original content through statistical characteristics. In contrast, the histogram of the final meaningful ciphertext closely aligns with that of the carrier image, showing high consistency in distribution shape, peak locations, and value ranges. This suggests that the embedding process introduces only minimal modifications to the carrier image, preserving its natural visual appearance while enabling covert transmission with high visual quality. Furthermore, the histogram of the decrypted image remains highly consistent with that of the plaintext, with nearly identical pixel value distributions, demonstrating that the decryption process accurately recovers the original information without introducing noticeable distortion or deviation. This statistical evidence further validates the reliability and reconstruction accuracy of the proposed method throughout the encryption and decryption process. Correlation Natural images exhibit strong correlations between adjacent pixels and among RGB channels. Noise-like encryption should suppress these correlations, while meaningful ciphertext should preserve correlations comparable to those of the cover image. In this work, the adjacent pixels correlation in three directions and the inter-channel correlation of the encrypted and decrypted results are shown in Figs. 17 and 18.Fig. 17Adjacent pixels correlation.Fig. 18Inter-channel correlation. Adjacent pixels correlation. Inter-channel correlation. As shown in Figs. 17 and 18, the correlation distribution of adjacent pixels in the plaintext is highly concentrated near the diagonal, exhibiting typical linear aggregation characteristics, which indicates strong structural dependency and local similarity between adjacent pixels in the original image. In contrast, after encryption, the noise-like ciphertext exhibits a more uniform and dispersed correlation distribution among adjacent pixels, no longer concentrated along the diagonal but evenly scattered across the entire coordinate space. This significant change confirms that the proposed scheme effectively disrupts the inherent pixel correlations within the image, eliminating statistical dependencies between adjacent pixels, thereby successfully concealing the structural information of the image and preventing attackers from obtaining effective clues through correlation analysis. Meanwhile, the correlation characteristics of the visually meaningful ciphertext are highly consistent with those of the carrier image, showing clear similarity in distribution patterns, aggregation degree, and scatter trends. This indicates that, while ensuring visual concealment, the ciphertext does not introduce detectable anomalies through statistical means, satisfying the stringent imperceptibility requirements of covert communication. Furthermore, the correlation distribution of the decrypted image is almost identical to that of the plaintext, with the scatter points once again tightly clustered along the diagonal, demonstrating that the decryption process accurately restores the structural relationships and pixel dependencies of the original image, further validating the effective preservation of reconstruction quality from a statistical perspective. Plaintext sensitivity analysis For noise-like ciphertext encryption schemes, the ciphertext must exhibit high sensitivity to slight changes in the plaintext to resist differential attacks. For meaningful ciphertext encryption, however, excessive sensitivity is undesirable, because the meaningful ciphertext is expected to remain visually consistent with the cover image; otherwise, the visual quality may be degraded. Table 4 reports how slight modifications to the plaintext influence the corresponding ciphertexts, quantified using the Number of Pixels Change Rate (NPCR) and the Structural Similarity (SSIM).Table 4Sensitivity analysis of plaintext.ModificationNoise-like ciphertextMeaningful ciphertextNPCRSSIMNPCRSSIM(1,1,1)+0.97830.00470.61450.9987(1,1,1)↔(512,512,1)0.97840.00860.61630.9987Encryption time0.97850.00920.61460.9987 Table 4Sensitivity analysis of plaintext.ModificationNoise-like ciphertextMeaningful ciphertextNPCRSSIMNPCRSSIM(1,1,1)+0.97830.00470.61450.9987(1,1,1)↔(512,512,1)0.97840.00860.61630.9987Encryption time0.97850.00920.61460.9987 Sensitivity analysis of plaintext. In Table 4, (1,1,1)+ denote increasing the pixel value at (1, 1, 1) by 1. The notation (1,1,1)↔ (512,512,1) represents swapping the pixel values at these two positions. From the experimental results, it can be observed that for noise-like ciphertext, the NPCR value approaches 1 before and after plaintext modification, while the SSIM value is close to 0. This indicates that even an extremely minor change in the plaintext leads to drastic and unpredictable variations in the ciphertext at the pixel level, demonstrating the high plaintext sensitivity of the proposed encryption algorithm. Such high sensitivity makes it difficult for attackers to infer the mapping relationship between plaintext and ciphertext through differential input, thereby effectively resisting differential attacks and ensuring the security of the encryption process. For meaningful ciphertext, although the NPCR value reaches approximately 0.6, indicating that changes in the plaintext still cause a certain degree of pixel value variation in the ciphertext, the SSIM value remains very close to 1, suggesting that such variations do not compromise the structural similarity between the ciphertext and the carrier image. From the perspective of visual perception, the meaningful ciphertext is nearly indistinguishable from the carrier image, maintaining a high level of visual consistency. Therefore, while ensuring encryption security, the meaningful ciphertext effectively achieves imperceptibility, enabling strong concealment capability during actual transmission. Chosen-plaintext and chosen-ciphertext attack analyses Chosen-plaintext attacks (CPA) and chosen-ciphertext attacks (CCA) are regarded as some of the most challenging threats to encryption systems. Therefore, this subsection evaluates the resistance of the proposed scheme against CPA and CCA to demonstrate its security. In the experiments, two types of attack images were considered: (i) slightly modified versions of the original plaintext (consistent with the plaintext sensitivity analysis), and (ii) the original plaintext itself but encrypted at different time instants. The dynamic keys derived from these attack images were then used to decrypt the original ciphertext. The resulting decrypted images are shown in Fig. 19.Fig. 19Chosen-plaintext attack. The first row shows the chosen attack images, and the second row shows the decrypted images obtained using the corresponding derived keys. Chosen-plaintext attack. The first row shows the chosen attack images, and the second row shows the decrypted images obtained using the corresponding derived keys. As illustrated in Fig. 19, even when the attack images are identical or highly similar to the original plaintext, the decrypted results obtained using the adversary-derived keys exhibit noise-like patterns from which no meaningful plaintext information can be recovered. This indicates that the proposed encryption system can effectively resist both CPA and CCA, preventing attackers from reconstructing valid keys based on additional plaintext or ciphertext samples. The resistance mainly stems from the introduction of plaintext hash values and current-time hash values during dynamic key generation, which grants the system a one-time-one-key property. Even when the plaintext remains unchanged, the keys obtained at different time instants are distinct, significantly enhancing security under CPA and CCA scenarios and ensuring both unpredictability of the encryption process and non-reusability of the key space. Comparative analysis To further illustrate the security of the proposed algorithm, Table 5 presents the information entropy of both the noise-like ciphertext and the visually meaningful ciphertext under different algorithms, as well as the NPCR and SSIM when the first pixel of the plaintext is increased by 1. In the table, the second column, denoted as “P & CO”, represents the plaintext and the cover image.Table 5Comparison of conventional security.RefP & CORef15Ref27Ref28Ref24Ref57OurNCInformation entropy6.15585.96056.65765.64675.99987.99836.5401NPCR–0.98230.98590.96390.98460.99620.9783SSIM–0.0186-0.00210.11730.07990.01020.0040CInformation entropy7.18817.29987.30167.40877.76577.40827.2773NPCR–0.67920.67680.69880.69820.90350.6145SSIM–0.99770.99810.99400.97680.93060.9987Bold indicates the best result, italics indicate the suboptimal result. Table 5Comparison of conventional security.RefP & CORef15Ref27Ref28Ref24Ref57OurNCInformation entropy6.15585.96056.65765.64675.99987.99836.5401NPCR–0.98230.98590.96390.98460.99620.9783SSIM–0.0186-0.00210.11730.07990.01020.0040CInformation entropy7.18817.29987.30167.40877.76577.40827.2773NPCR–0.67920.67680.69880.69820.90350.6145SSIM–0.99770.99810.99400.97680.93060.9987Bold indicates the best result, italics indicate the suboptimal result. Comparison of conventional security. Bold indicates the best result, italics indicate the suboptimal result. From Table 5, it can be seen that the noise-like ciphertext generated by the proposed algorithm effectively destroys the plaintext information, with information entropy increasing from 6.1558 to 6.5401. Although this is lower than that of57, it is significantly better than those of15,24,28. This is mainly because57 employs a secondary diffusion operation, resulting in a more uniformly distributed noise-like ciphertext. In contrast, the proposed algorithm and those in15,27,28 assume that the compression process already performs one diffusion operation, making a second diffusion unnecessary. However, to improve the uniformity of the noise-like ciphertext, the bit scrambling adopted in this paper not only changes pixel positions but also modifies the measurement values. Compared with15 and28, it can also be seen that the bit scrambling used in this paper and in27 is effective. Since no secondary diffusion operation is performed, the proposed algorithm, like those in15,27,28, uses the SHA value of the plaintext as part of the dynamic key generation, leveraging the sensitivity of SHA functions and chaos to enhance plaintext sensitivity. This achieves a one-time one-key encryption effect, providing strong resistance against chosen-plaintext and chosen-ciphertext attacks, as well as differential attacks. As shown in Table 5, under a slight change in the plaintext, the NPCR of the noise-like ciphertext is close to 1, and the SSIM between the ciphertexts is close to 0. This indicates that the two ciphertexts are completely different, demonstrating strong resistance to differential attacks. For the meaningful ciphertext, Table 5 shows that the information entropy of the proposed algorithm is closer to that of the cover image. The NPCR of the ciphertext before and after plaintext modification are the lowest, and the structural similarity is closer to 1. This indicates that regardless of plaintext changes, the meaningful ciphertext remains closer to the cover image with smaller variations, indirectly demonstrating that the proposed algorithm achieves better ciphertext quality. Encryption and decryption efficiency comparison Table 6 compares the average encryption and decryption times of different algorithms. The reported results are based on the average encryption and decryption times of the four plaintext-ciphertext image pairs shown in Fig. 13. All experiments were conducted on a personal computer equipped with 64 GB of RAM, an Intel i7 processor, and the Windows 11 operating system, using MATLAB R2018a as the simulation platform. For24 and57, deep learning-based compression and reconstruction were performed using the CPU, without GPU acceleration. To facilitate the explanation of time complexity, let the plaintext size be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m \times n \times 3$$\end{document} and the cover image size be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M \times N \times 3$$\end{document}.Table 6Average encryption and decryption time.RefRef15Ref27Ref28Ref24Ref57OurCompression encryption0.52560.54670.45190.30483.95300.8111Embedded0.35280.12661.05720.25290.04260.0996Extract0.27440.08640.39120.06420.03360.0700Decryption reconstruction0.76330.66394.77130.26240.33940.7241Total time1.91611.42366.67160.88434.36861.7048Bold indicates the best result, italics indicate the suboptimal result. Table 6Average encryption and decryption time.RefRef15Ref27Ref28Ref24Ref57OurCompression encryption0.52560.54670.45190.30483.95300.8111Embedded0.35280.12661.05720.25290.04260.0996Extract0.27440.08640.39120.06420.03360.0700Decryption reconstruction0.76330.66394.77130.26240.33940.7241Total time1.91611.42366.67160.88434.36861.7048Bold indicates the best result, italics indicate the suboptimal result. Average encryption and decryption time. Bold indicates the best result, italics indicate the suboptimal result. As shown in Table 6, the proposed method is significantly faster than those in28 and57. This is because the reconstruction in28 requires multiple iterations, while57 employs a meta-learning strategy that iteratively updates the measurement matrix during the sampling process, involving repeated iterations and expensive parameter updates. In contrast, the proposed method assigns higher sampling weights to low-frequency components in the upper-left region of the sparse plaintext, thereby requiring only a few iterations to achieve satisfactory reconstruction and allowing early termination of the reconstruction process. In the sampling stage, the proposed method adopts a stairwise sampling strategy. The complexities for the three channels are \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(0.96m^2n + 0.9216mn^2),$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(0.36m^2n + 0.1296mn^2),$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(0.18m^2n + 0.0324mn^2),$$\end{document} respectively, resulting in a total complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(1.5m^2n + 1.0836mn^2).$$\end{document} In contrast,15 and27 employ the same two-dimensional compressive sampling across all three channels, with a complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(1.5m^2n + 0.75mn^2).$$\end{document} Shiwei28 performs one-dimensional column-wise compression with a complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(0.75m^2n),$$\end{document} which is lower. During encryption, although the non-repetitive global scrambling used in28 and15 has a complexity of O(3mn), which is higher than the bit-level row-column scrambling of O(3m+8n) adopted in this paper and the M-ary decomposition-based bit scrambling of O(3m+4n) used in27, these complexities are negligible compared to the sampling stage. Therefore28, achieves lower compression and encryption times than the proposed method and those in15,27. Chen et al.24 achieves the lowest encryption time due to the absence of a scrambling process. In the embedding and extraction stages28, requires searching for appropriate histogram shifting parameters and performing multiple embeddings to obtain the ciphertext with optimal PSNR, making it the most time-consuming approach, while the other methods require relatively little time. Decryption and reconstruction are the most time-consuming stages for the proposed method and15,27,28, mainly because the reconstruction algorithm requires multiple iterations, each involving two matrix multiplication processes similar to those in the compression stage, resulting in high complexity that is further amplified by the multiple iterations. In contrast24 and57, require only a few convolutions after coarse reconstruction, and when the convolution kernel is 3 × 3, the complexity is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(c_1 c_2 9mn),$$\end{document} enabling faster reconstruction, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c_2$$\end{document} denote the input and output channel dimensions, respectively. Although the proposed method does not achieve the fastest running speed, its efficiency can be further improved by modifying the scrambling, compression, and reconstruction strategies. For instance, adopting pixel-level scrambling instead of bit-level scrambling, or omitting scrambling altogether as in24 and57, could further increase speed. For compressive sensing, a deep learning-based compressive sensing approach such as that in24 can be adopted for each channel while maintaining stairwise sampling, preserving good reconstruction quality while achieving fast compression and reconstruction. Overall, the proposed scheme achieves a total encryption and decryption time of 1.7048 seconds, which is lower than those reported in15,28,57. This result indicates that the proposed encryption framework maintains relatively high computational efficiency. The main computational cost of the proposed scheme lies in the generation of chaotic sequences, which cannot be parallelized, and the image compression and reconstruction stage, which has high time complexity. Future work will focus on optimizing these two processes to improve overall system efficiency. Ciphertext quality comparison To compare the ciphertext quality of different algorithms, Table 7 presents the comparison results of PSNR, SSIM, NPCR, and UACI (Unified Average Changing Intensity) between the ciphertext and the cover image. Figure 20 also provides the visualization results of the ciphertext for two cases of plaintext and cover images.Table 7Ciphertext quality.PlaintextCover imageIndicatorsRef15Ref27Ref28Ref24Ref57OurLeanBaboonPSNR(dB)46.369348.128648.432442.182831.071950.1782SSIM0.99900.99920.99550.99150.94570.9996NPCR0.75030.55070.58840.56180.99990.4363UACI0.00390.00270.00280.00510.49390.0020ToucanPortofinoPSNR(dB)46.369447.924248.343042.175432.945749.7839SSIM0.99830.99890.99260.98540.90180.9993NPCR0.74970.57830.57540.56271.00000.4673UACI0.00390.00290.00260.00510.38450.0021Butfish1BarnfallPSNR(dB)46.384847.984148.116642.178434.813750.4520SSIM0.99760.99830.99830.98730.93950.9991NPCR0.74300.56690.60910.56250.99950.4319UACI0.00390.00280.00280.00510.30490.0019ElephantSailboatPSNR(dB)46.358247.931648.082242.183634.143050.0912SSIM0.99770.91850.98010.98060.90920.9992NPCR0.75020.58770.57280.56260.99840.4411UACI0.00390.00290.00270.00510.48200.0020MeanPSNR(dB)46.370447.992148.243542.180133.243650.1263SSIM0.99820.97870.99160.98620.92410.9993NPCR0.74830.57090.58640.56240.99940.4441UACI0.00390.00280.00270.00510.41630.0020Bold indicates the best result, italics indicate the suboptimal result. Table 7Ciphertext quality.PlaintextCover imageIndicatorsRef15Ref27Ref28Ref24Ref57OurLeanBaboonPSNR(dB)46.369348.128648.432442.182831.071950.1782SSIM0.99900.99920.99550.99150.94570.9996NPCR0.75030.55070.58840.56180.99990.4363UACI0.00390.00270.00280.00510.49390.0020ToucanPortofinoPSNR(dB)46.369447.924248.343042.175432.945749.7839SSIM0.99830.99890.99260.98540.90180.9993NPCR0.74970.57830.57540.56271.00000.4673UACI0.00390.00290.00260.00510.38450.0021Butfish1BarnfallPSNR(dB)46.384847.984148.116642.178434.813750.4520SSIM0.99760.99830.99830.98730.93950.9991NPCR0.74300.56690.60910.56250.99950.4319UACI0.00390.00280.00280.00510.30490.0019ElephantSailboatPSNR(dB)46.358247.931648.082242.183634.143050.0912SSIM0.99770.91850.98010.98060.90920.9992NPCR0.75020.58770.57280.56260.99840.4411UACI0.00390.00290.00270.00510.48200.0020MeanPSNR(dB)46.370447.992148.243542.180133.243650.1263SSIM0.99820.97870.99160.98620.92410.9993NPCR0.74830.57090.58640.56240.99940.4441UACI0.00390.00280.00270.00510.41630.0020Bold indicates the best result, italics indicate the suboptimal result. Ciphertext quality. Bold indicates the best result, italics indicate the suboptimal result. Fig. 20Ciphertexts of different algorithms. The first and third rows display the corresponding ciphertext images, whereas the second and fourth rows show the difference maps between the red-boxed ciphertext regions and the cover images, magnified 30×. Ciphertexts of different algorithms. The first and third rows display the corresponding ciphertext images, whereas the second and fourth rows show the difference maps between the red-boxed ciphertext regions and the cover images, magnified 30×. As shown in Table 7, the proposed method achieves optimal performance across all metrics, significantly outperforming the other algorithms. Improvements of 1.8828 dB, 0.0011, 0.1183, and 0.0007 are achieved in PSNR, SSIM, NPCR, and UACI, respectively, demonstrating that the proposed algorithm produces ciphertext with high quality. Among the compared methods27,28, outperform15,24, primarily because they employ histogram shifting to reduce the overall magnitude of the noise-like ciphertext, resulting in smaller modifications to the ciphertext relative to the cover image and lower change intensity. This is further evidenced by the reduction in NPCR and UACI from 0.75 and 0.004 to 0.58 and 0.0028, respectively. In contrast, the proposed method adopts histogram rearrangement to further reduce the number of 0-bits in the noise-like ciphertext, combined with embedding value adjustment, which leads to even smaller modifications to the cover image, achieving NPCR and UACI values of 0.44 and 0.002, respectively. These numerical results demonstrate the effectiveness and advancement of the proposed method. The difference images between the ciphertext and the cover image in Fig. 20 show that the proposed method produces significantly darker results, further confirming that the visual distortion introduced to the cover image is extremely small. Overall, the proposed algorithm provides high-quality ciphertext that is much closer to the cover image, effectively ensuring imperceptibility and security during ciphertext transmission. Reconstructed image quality comparison Table 8 reports the quantitative results of reconstructed image quality obtained by different methods, and Fig. 21 illustrates the visual reconstruction results for two test images.Table 8Reconstructed image quality.Plaintextcover imagesRef15Ref27Ref28Ref24Ref57OurLeanBaboon32.8809/0.985233.4185/0.986832.3737/0.810934.0164/0.892532.9582/0.854734.4580/0.9902ToucanPortofino37.7422/0.770939.2235/0.778037.4694/0.895440.3417/0.960836.5425/0.843840.5288/0.9709Butfish1Barnfall27.0466/0.781127.2046/0.785625.8938/0.685626.9646/0.809026.8806/0.760628.7658/0.9218ElephantSailboat32.3468/0.911232.6587/0.918530.9262/0.792432.8874/0.874931.6389/0.843133.4474/0.9554Mean32.5041/0.862133.1263/0.867231.6658/0.796133.5525/0.884332.0050/0.825634.3000/0.9596Bold indicates the best result, italics indicate the suboptimal result. Table 8Reconstructed image quality.Plaintextcover imagesRef15Ref27Ref28Ref24Ref57OurLeanBaboon32.8809/0.985233.4185/0.986832.3737/0.810934.0164/0.892532.9582/0.854734.4580/0.9902ToucanPortofino37.7422/0.770939.2235/0.778037.4694/0.895440.3417/0.960836.5425/0.843840.5288/0.9709Butfish1Barnfall27.0466/0.781127.2046/0.785625.8938/0.685626.9646/0.809026.8806/0.760628.7658/0.9218ElephantSailboat32.3468/0.911232.6587/0.918530.9262/0.792432.8874/0.874931.6389/0.843133.4474/0.9554Mean32.5041/0.862133.1263/0.867231.6658/0.796133.5525/0.884332.0050/0.825634.3000/0.9596Bold indicates the best result, italics indicate the suboptimal result. Reconstructed image quality. Bold indicates the best result, italics indicate the suboptimal result. Fig. 21Reconstructed images of different algorithms. The second and fourth rows show the reconstructed results corresponding to the red-boxed regions in the first and second rows. Reconstructed images of different algorithms. The second and fourth rows show the reconstructed results corresponding to the red-boxed regions in the first and second rows. From Table 8, it can be observed that the proposed method achieves the best reconstruction performance in terms of both PSNR and SSIM, outperforming the competing schemes by nearly 1 dB in PSNR and approximately 0.1 in SSIM on average. As shown in Fig. 21, the reconstructed images produced by the proposed method are visually clearer and preserve more structural details. The superior reconstruction quality mainly stems from the use of 3D-DCT based sparsification combined with a stairwise compression strategy. This design effectively exploits the inter-channel redundancy of color images, which is largely ignored by other methods. Consequently, under the same sampling rate, the obtained measurements retain more informative components, ensuring high-quality image reconstruction. Discussion The above comparisons indicate that the proposed scheme achieves a well-balanced performance in terms of ciphertext imperceptibility, reconstruction quality, and computational efficiency. Benefiting from histogram shifting and embedding value adjustment, the generated ciphertext introduces minimal distortion to the cover image, resulting in superior PSNR and SSIM and ensuring strong visual imperceptibility. Moreover, the integration of 3D-DCT sparsification with stairwise compression effectively exploits inter-channel redundancy in color images, enabling more informative measurements to be preserved under the same sampling rate. This leads to consistently improved reconstruction quality compared with existing methods. In terms of efficiency, the proposed method avoids excessive iterative reconstruction by prioritizing low-frequency components and allowing early termination, yielding competitive encryption and decryption speed without relying on complex learning models. Conclusion This paper proposes a novel meaningful ciphertext encryption scheme. The proposed framework first employs a two-dimensional CICCM hyperchaotic system to generate control sequences for encryption, followed by stairwise compressive sensing applied to the plaintext image, and finally embeds the resulting ciphertext into a cover image. Compared with conventional two-dimensional chaotic systems, the proposed CICCM system exhibits superior chaotic characteristics, particularly in terms of the maximum Lyapunov exponent and sequence distribution, demonstrating its suitability for secure encryption applications. During the compression stage, a 3D-DCT is adopted to exploit the energy attenuation characteristics across RGB channels. Furthermore, a stairwise sampling strategy is introduced, where low-frequency components are assigned higher sampling rates. This strategy simultaneously reduces spatial redundancy and inter-channel redundancy, thereby improving the reconstruction quality of decrypted images. In addition, a histogram reorganisation and embedded-value adjustment strategy is designed to adapt to the approximately Gaussian distribution of compressive sensing measurements. This approach effectively reduces distortion to the cover image and improves ciphertext visual quality. Comparative experimental results demonstrate that the proposed scheme achieves high-quality ciphertext generation, superior reconstruction performance, fast encryption speed, and strong security robustness.