BPBiLSTM-IDS: a lightweight intrusion detection framework for cyber-physical UAV networks Authors: Hafiz Muhammad Attaullah, Inam Ullah Khan, Muhammad Mansoor Alam, Mazliham Mohd Su’ud, Keshav Kaushik, Ahthasham Sajid, Nurashikin Saaludin, Talha Ahmed Khan Published: 2026 ABSTRACT AbstractUnmanned Aerial Vehicles (UAVs) have revolutionized modern technology by enabling autonomous operations in dynamic environments; however, their reliance on wireless networks exposes them to significant cybersecurity threats. These threats include De-authentication Denial of Service, False Data Injection (FDI), Replay, and Evil Twin attacks, which severely impact usability and data integrity. Conventional Intrusion Detection Systems (IDS) suffer from drawbacks such as high false alarm rates, excessive resource consumption, and non-proportional mechanisms for dynamic UAV topologies. To address these challenges, this study introduces an enhanced AIDS architecture in which optimal features are selected using Binary Pigeon Optimization (BP), and intrusion detection is performed using a Bidirectional Long Short-Term Memory (Bi-LSTM) with 1D-CNN model. BP enables feature selection independent of computational cost, mitigating the impact of high-cost or exhaustive features, while Bi-LSTM effectively captures temporal characteristics of UAV network traffic for accurate attack detection. Experimental evaluation on a cyber-physical UAV dataset demonstrates that the proposed BP + Bi-LSTM model along with 1D-CNN outperforms traditional ML approaches such as Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), and Deep Neural Network (DNN), achieving an accuracy of 98.74% ± 0.07 (mean ± std over 10 runs), along with high precision, recall, and an optimal false positive rate. These results confirm that the proposed model is a scalable, adaptive, and lightweight solution for real-time intrusion detection in UAV networks.Keywords: Unmanned aerial vehicles, Intrusion detection system, Anomaly detection, Binary pigeon optimization, Bi-LSTM, Cyber-physical security, UAV-IDSSubject terms: Engineering, Mathematics and computing Introduction In recent years, Unmanned Aerial Vehicles (UAVs) have assumed significant roles across a wide range of sectors, such as military operations, administrative services, and even commercial applications, including surveillance, transportation, agriculture, climate monitoring, and rescue missions. UAV networks have enhanced operational efficiency through real time data collection and remote control features as well as capabilities of real time data collection and remote control features as well as real time data collection and remote control capabilities1. Nonetheless, since these networks are based on the idea of heavy wireless communication between UAVs and ground control stations, as well as cloud services, they also have a lot of vulnerabilities. Therefore, they are often the victims of high-level cyber threats, among which, in particular, are spoofing, False Data Injection (FDI), and Denial of Service (DoS) attacks. These intrusions may result in disastrous loss of control, gross data leakage or actual crashing of UAVs2,3. Due to the resource-limited and highly dynamic topology of the UAV networks, the challenge of ensuring the security of such communication channels without exceeding the onboard computing capacity is a pressing issue. As shown in Fig. 1, UAV networks have six broad categories of connection (A1-A6) and node types that, incidentally, correspond to potential vectors of threat (re-produced with reference4). The above figure provides an architecture that provides a broad background of UAV and vulnerability and the importance of proactive mechanisms to secure the UAV networks. Illustration of communication pathways between UAVs, ground devices, and cloud services with security threat vectors4. The UAV networks are not like the traditional networks because they have a highly dynamic topology, mobility of nodes and limited resource constraints. Intrusion Detection Systems (IDS) are therefore critical in the monitoring of network traffic and detection of malicious traffic. IDS techniques are generally classified into signature-based and anomaly-based approaches. IDS based on misuse are based on pre-determined attack signatures and cannot be used to detect unknown or emerging threats5. On the other hand, conventional AIDS are usually characterized by high false positive, reduced detection accuracy, and high computational costs6. To overcome these limitations, advanced IDS frameworks that improve detection accuracy while reducing computational complexity and adapting to UAV network requirements are essential. The recent studies focus on the application of smart classification methods, such as machine learning (ML) and deep learning (DL) algorithms, to improve the performance of intrusion detection methods4,7. These classifiers are normally judged on the basis of detection rate, false alarm rate and the overall accuracy. The use of selective ML and DL techniques has been extensively used to detect intrusions in network environments with high accuracy8–10. Recent advancements reflect a strong trend toward integrating explainable AI (XAI), attention mechanisms, and transfer learning to secure resource-constrained IoT and Cyber-Physical Systems (CPS). For instance, scaled dot-product attention has been utilized for privacy-preserving anomaly detection in Industrial IoT11, while autonomous LLM agents with memory-augmented, edge-optimized architectures demonstrate robust zero-day resilience12. Similarly, explainable ensemble transfer learning approaches have been highly effective for botnet attack detection in dynamic vehicle networks13. Furthermore, optimized architectures, such as modified genetic algorithms combined with edge-capable LSTM networks14 and Autoencoder-driven DCNN-LSTM models15, have set new benchmarks for precision in critical infrastructure. While these models excel in general IoT or industrial settings, adapting such high-performance logic specifically to the strict aerodynamic and energy constraints of UAVs requires further optimization. The typical ML-based IDS methods are Support Vector Machine (SVM), Logistic Regression (LR), Naïve Bayes (NB), Artificial Neural Networks (ANN), and Dimensionality Reduction (DiR) methods16–18. Nevertheless, the classical models have problems including high false alarm rates, complexity of modeling, reduced detection accuracy, and high dimensional feature space19,20. Such constraints encourage the creation of a new IDS model that is specific to UAV networks. The paper presents an improved UAV-IDS model, which incorporates Binary Pigeon Optimization (BP) for feature selection and a Bi-LSTM with 1D-CNN to identify anomalies, extending the baseline in21 from a general cybersecurity context to a UAV-specific context. On the basis of the behavior of foraging in pigeons and then 1D-CNN, BP discovers an optimal set of features, dimensionality reduction and enhanced classification performance and accuracy. The model of Bi-LSTM operates with sequential data in both directions, i.e., forward and backward directions, and therefore, it can be successfully used to capture the temporal dependencies and complex traffic patterns in UAV networks. This enables a great level of differentiation between regular and malicious activities and reduces false positives. The key contributions of this paper can be summarized as follows: Optimized Feature Selection: BP is employed to select an optimal feature subset, enhancing detection accuracy and reducing computational overhead. Optimized Feature Selection: BP is employed to select an optimal feature subset, enhancing detection accuracy and reducing computational overhead. Bi-LSTM and 1D-CNN-Based Classification: The optimized features are used to train a Bi-LSTM model capable of effectively classifying normal and attack traffic by leveraging temporal dependencies. Bi-LSTM and 1D-CNN-Based Classification: The optimized features are used to train a Bi-LSTM model capable of effectively classifying normal and attack traffic by leveraging temporal dependencies. Performance Evaluation: Extensive experiments demonstrate the superiority of the proposed framework over conventional approaches in terms of detection rate, false alarm rate, and processing time. Performance Evaluation: Extensive experiments demonstrate the superiority of the proposed framework over conventional approaches in terms of detection rate, false alarm rate, and processing time. Scalability and Real-Time Validation: The framework is validated for scalability and real-time applicability using comprehensive UAV cyber-physical datasets. Scalability and Real-Time Validation: The framework is validated for scalability and real-time applicability using comprehensive UAV cyber-physical datasets. The remainder of this paper is organized as follows. Section "UAV-based IDS approaches" provides a literature review and determines drawbacks of current IDS solutions to UAV networks. Section “BPBiLSTM optimization” shows the elaborated BP-BiLSTM-based framework, which consists of its elements of the methodology. Section “Simulation environment” outlines the simulation algorithm, implementation set-up and evaluation procedures. Section "Results and discussions" expounds on the datasets used and make a comparative analysis with the existing approaches. Section “Conclusion” covers the experimental findings and the lessons obtained out of the analysis. Lastly, a conclusion is provided in Section 7 where the study ends with a short future research directions. Related work This section presents a comprehensive review of existing IDS methodologies proposed to enhance the security of UAV networks. The studies reviewed are discussed in the context of their effectiveness, benefits, and shortcomings in identifying and reducing cyber threats in UAV settings. In22, the authors implemented a neuro-fuzzy IDS system to enhance the security of UAV networks against cyber attacks. The proposed centralized solution was aimed at the effective transfer of reliable data and the necessity of lightweight security solutions to maintain UAV maneuverability. The paper has emphasized the necessity of flexible IDS models because of the topology that is very dynamic and the environmental conditions of UAV networks are different. The Authors in23, proposed a cross-layer IDS of UAV networks using an Anomaly-based Model of Features (AMoF) to enhance the accuracy of anomaly detection. The framework improved detection accuracy by studying the UAV behavior at various protocol layers. The cross-layer design is however very computationally expensive and thus not as applicable in resource constrained UAV platforms. A thorough review of ML and DL-based IDS methods in UAVs was presented in24, aiming to enhance the accuracy of attack classification and minimize the misclassification error. IDS approaches are generally categorized based on detection mechanisms, including anomaly-based, misuse-based, clustering-based, hybrid-based, trust-based, and zone-based IDS. Since UAV networks are dynamic in topology and mobile, the IDS frameworks should be energy-efficient, precise, and scalable. In25, an IDS based on binary optimization was proposed, in which feature selection was important to attain a higher accuracy with the lower dimensionality. Nevertheless, the research indicated that the processing demands were high, which restricts the applicability of the study to real-time in UAV contexts. In26, an improved Multi-Kernel Extreme Learning Machine (MK-ELM) was proposed to increase the detection accuracy and Quality of Service (QoS) in UAV networks. Although the multi-kernel processing was more accurate, it also increased the computational complexity making it less appropriate in time-critical UAV applications. Equally30, presented an IDS model using Slime Mould Algorithm (SMA) to enhance reliability whereas31 concentrated on minimizing false alarms. Despite their effectiveness, these models have challenges in their practical implementation because of the UAV computational limitations. In27, a lightweight IDS was proposed that is based on K-Nearest Neighbors (KNN) and Sine Cosine Algorithm (SCA) to minimize the false alarm rates and maximize the detection accuracy. Although it can be used in low-resource settings, the framework continues to be inefficient in feature selection and requires a lot of time to train. In the article, a Deep Neural Network (DNN)-based IDS that overcomes the problem of class imbalance was introduced in32 and cross-correlation-based feature selection was used to enhance the accuracy. Also, the reference33 used a conditional Generative Adversarial Network (GAN) to create realistic traffic data, but the detection accuracy and the quality of service were poor. An AI-based anomaly detection model using Decision Tree (DT) and Naïve Bayes enhanced with ANN was introduced in34. The performance measures considered in the study included runtime and memory consumption, although a lot of parameter optimization was necessary. A CNN–LSTM-based IDS was proposed in28 to improve detection accuracy; however, its high computational complexity limits UAV deployment feasibility. Furthermore, a three-tier data mining-based IDS framework was introduced in35, employing ANN, KNN, and SVM classifiers for intrusion detection. Similarly29, utilized a CNN–LSTM model with emphasis on data encoding and normalization. Despite improved detection rates, both approaches suffer from high training times and complex system architectures. To better identify research gaps, Table 1 summarizes the limitations of existing IDS methodologies for UAV networks. Table 1Summary of limitations in existing IDS methodologies for UAV networks.Ref.Proposed MethodologyDatasetParametersIDS TypeKey Limitations22Neuro-fuzzy IDSNSL-KDDFuzzy rules, membership functionsHybridCentralized design limits scalability; poor adaptability to dynamic UAV environments.23Cross-layer IDS with AMoFCICIDSPacket flow features, anomaly thresholdsHybridMemory-intensive multi-layer feature mapping restricts edge deployment.24Hybrid ML/DL IDS comparison–Feature selection, DL architecturesHybridLacks real-world UAV dataset validation.25Binary optimization-based IDSBoT-IoTFeature ranking, binary classifiersSignatureHigh processing cost limits real-time UAV deployment.26MK-ELM IDSCSE-CIC-IDSKernel optimization, weight tuningHybridKernel weight tuning introduces severe scalability bottlenecks.27KNN + SCA IDSAWIDk-value, similarity measuresAnomalyInefficient feature selection; long training time.28CNN–LSTM IDSTON_IoTLayer depth, sequence lengthAnomalyPropensity for overfitting in minority classes during sequence learning.29Three-tier data mining IDSUAV datasetClustering thresholds, dimensionality reductionHybridStrict dependency on static clustering thresholds limits dynamic adaptability. Summary of limitations in existing IDS methodologies for UAV networks. Recent literature has increasingly explored hybrid swarm-intelligence and sequential deep learning models for anomaly detection. For instance, hybrid optimization techniques have been successfully applied to neural networks for DDoS detection36,37, and feature-optimized models have shown robust success in complex, dynamic prediction tasks38–40. However, existing hybrid IDS models often struggle to balance the exhaustive search space of feature selection with the sequential complexity of deep architectures. The novelty of our proposed BP + Bi-LSTM approach lies in its mathematically lightweight map-compass operator, which uniquely bounds computational complexity to O(N · D). Unlike traditional swarm methods that suffer from premature convergence or high memory overhead, BP dynamically selects optimal features without stalling edge-deployable hardware, creating a highly synergized pipeline specifically tailored for UAV aerodynamic and battery constraints. Regardless of the improvements in the IDS frameworks, it is important to mention that the limitations are still critical to overcome the issue of the UAV networks and attain an effective and reliable intrusion detection. Based on the reviewed IDS frameworks, the objectives were focused on the detection rates of increases, reduced false positive, and learning optimization. Nevertheless, there are still a number of challenges, like High Training and Testing Time: In the vast majority of cases, the IDS frameworks include the complex training and testing phases; hence, the process of real-time detection of anomaly is complicated. Complex Feature Selection: Feature extraction is often a time-consuming process, making computational efficiency a significant challenge. Oversampling Issues: To balance the data, some IDS models rely on oversampling. While this improves overall detection, it frequently increases the false positive rate. Also, there is reliability limitations. The issue is in not all the IDS models have adequate reliability in dynamic UAV environment and, specifically, high mobility operation. Implementation Complexity: The heavy computational demands of certain models make deployment on resource-constrained UAVs highly challenging. UAV-based IDS approaches Several approaches to IDS have been suggested to efficiently identify intrusions in UAV networks, as illustrated in Fig. 2. Classification of IDS based on detection mechanisms. These approaches are generally categorized into signature-based, anomaly-based, and hybrid-based methods, each designed to address specific security challenges in UAV environments. Signature-based IDS for UAVs Signature-based IDS (SIDS) is a type of threat detector that matches network traffic with a pre-established database of known attack signatures26. The technique is highly efficient in identifying well-documented attacks and maintains a low false-positive rate. It does not however detect zero-day attacks, and new threats, which limits its use in UAV networks where new threats are the common occurrence. In the context of UAVs, SIDS can be best used to detect recurrent attack patterns such as GPS jamming or spoofing. As an example, a SIDS can automatically identify malicious commands that have been previously registered against control systems of UAVs. Nevertheless, they depend on a fixed set of signature databases, which restricts their capacity to identify a zero-day attack or a new variant of threat. This is a reason why this weakness is extremely important in UAV networks, where attackers often invent new ways of capitalizing on the vulnerabilities. Although these are the limitations, SIDS still play a crucial role in multi-layered UAV security models. Complementing AI-driven detection techniques or anomaly-based detection, signature-based systems offer extra protection against established threats. Anomaly-based IDS for UAVs Anomaly-based IDS (AIDS) model normal system behavior and flag deviations from this baseline as potential intrusions. This approach is well suited for identifying unknown or zero-day attacks, making it highly adaptive and scalable for dynamic UAV environments20. Unlike signature-based approaches, AIDS do not depend on predefined attack patterns. Such intrusions can be detected by AIDS which identifies abnormalities in the communication traffic such as unexpected peaks in data transmission or abnormal telemetry in the UAV or zero day malware. However, the systems that rely on anomalies do not lack problems. They are more likely to generate high false positive as compared to the signature based IDS. A threat can also be reported as any deviation of the baseline. This is an issue particularly with UAV networks where the environment or the operations being conducted may change causing valid deviations. To overcome these problems, new AI and ML developments have also been integrated into the anomaly detection to increase their flexibility and false alarms. Hybrid-based IDS for UAVs Hybrid IDS combine the strengths of both signature-based and anomaly-based approaches. It also uses the already known signatures to identify fast and simultaneously monitors the anomalies to identify the new threats. This dual approach provides higher precision and fewer false positives compared to standalone methods. Wang et al.27 emphasizes that the hybrid IDS structure is a moderate solution to UAV networks, both in terms of efficiency and accuracy. HIDS can be applied in the case of UAVs to monitor critical processes of firmware integrity, operating system processes, and file system changes. As an example, a HIDS can detect any malicious attempt to modify flight control software or install malicious code to a UAV. This is particularly true of military applications or high-risk applications, where UAVs are often the victims of advanced persistent threat (APTs) who would like to disrupt their operation or steal sensitive data. HIDS has a number of disadvantages when used in UAVs despite its benefits. These systems demand computational power to process logs and track processes which may demand the already constrained onboard capabilities of UAVs. Also, because HIDS can only give insight into the host system it adjoins, it cannot adequately detect attacks on the network or organized attacks in which more than one UAV is used. However, combined with network based IDS, HIDS provides an essential point of defense in that it safeguards on a host level, as well as a network level. Although AIDS techniques are strong in detecting new threats, they need improvement to minimize false positives and computation costs. The dual aspect of hybrid techniques seeks to strike a compromise between these concerns though caution should be taken in its application in order to preserve real-time detection abilities. BPBiLSTM IDS The primary contribution of this study is the development of a computationally intelligent IDS framework that will help in safeguarding UAV networks against cyber attacks. The suggested model combines an effective method of BP with a Bi-LSTM classifier, which is also known as the BPBiLSTM model. The standard feature selection methods used in the IDS research field and were used extensively include Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Recursive Feature Elimination (RFE)24,26,27. Nevertheless, according to the report on the same41, these methods tend to be slow in convergence and computationally intensive, and this is why they are not as resource-constrained UAV-applicable. The rationale behind the use of BP in the present study is that it is lightweight, converges faster, and it can be applied to work with binary feature space. The general working process of the suggested architecture is shown in Algorithm 1. It includes four major steps: data preparation, feature selection with the assistance of BP, Bi-LSTM-based classification, and performance evaluation. The cleaning, normalizing and clustering of the data to ensure that it gives balanced and quality input is the first step. The BP then chooses the optimal feature subset by the position and fitness of the features thereby reducing the cost of computation and maximizing the detection accuracy. The Bi-LSTM classifier is then trained with the chosen features and it makes use of temporal dependencies in the UAV network traffic to identify anomalies. Finally, the performance indicators are determined to verify the relevance of the suggested framework in the real time. Mathematically, BP stands out from high-cost binary optimizations (the conventional GA or complex variations of the PSO). Whereas standard GA involves calculating multi-point crossover and mutation permutations which are complex and have a per-generation complexity of (where N is population and D is dimensions), BP is based solely on binary position toggling through a simplified map-compass operator. This bounds the BP computational complexity to O(N · D) per iteration. BP can automatically reduce memory usage and CPU time, as it removes the constant maintenance of vectors of velocity, and the intricate heuristic resampling, which explains its classification as a very lightweight optimizer. BPBiLSTM Intrusion Detection Framework for UAV Networks The proposed framework is applied to a publicly available cyber-physical UAV dataset with normal operations and cyberattacks42–44. There are four large categories of attacks in the dataset, namely False Data Injection, De-authentication DoS, Replay, and Evil Twin attacks. Preprocessing of data involves cleaning, dataset splitting, normalization, and clustering to achieve reliability and balance before optimizing the features. To strictly prevent data leakage, the dataset is first partitioned into training and testing sets. Data normalization is then performed by fitting a Min-Max scaler exclusively on the training data to transform feature values into a comparable scale between 0 and 1. The exact same scaling parameters (minimum and maximum values) derived from the training set are subsequently applied to transform the test set. This ensures the model does not gain unauthorized access to the test distribution during training. The normalization function is expressed as: To ensure a balanced data distribution and manage computational overhead, we employ K-Means clustering in conjunction with the Synthetic Minority Over-sampling Technique (SMOTE). Specifically, the K-Means algorithm is applied with k=5, which is the number of major traffic classes that are available in the data set (Normal, De-authentication DoS, Replay, FDI and Evil Twin). In this distance-based approach samples are allocated to the centroids in an iterative process until the centroids converge. After the cluster assignments, SMOTE is used to make synthetic samples in the minority clusters in such a way that there is an equal representation of all the categories of attacks. This is a hybrid clustering-resampling technique, which ensures that the model does not favor majority classes but also does not randomly generate unproductive synthetic data to produce a highly representative and balanced dataset to use in further feature optimization. The BP fitness function is defined as: where α and β represent the weighting factors for accuracy and feature subset size, respectively. The update equations of both the velocity and position of BP make use of a special map and compass variable, denoted as γ, to direct the search in the optimization (not to be confused with the fitness weights): A sigmoid-based thresholding function ensures binary feature selection: where and θ = 0.5. The center pigeon position is calculated as: The classification phase employs a Bi-LSTM model that has the ability to learn long-term dependencies using data in both forward and backward directions. The Bi-LSTM model was thoroughly fine-tuned for classification accuracy and to avoid overfitting, using the Optuna hyperparameter tuning package. We performed a sensitivity analysis over a search space for hidden units (32 to 256), dropout (0.1 to 0.5) and learning rates ( to ). As the best-performing model based on the highest validation accuracy, the final Bi-LSTM model architecture consists of two layers of Bi-LSTM with 128 hidden units, a dropout rate of 0.3, and uses the Adam optimizer with a learning rate of 0.001. This is trained for 50 epochs with a batch size of 64 and employing binary cross-entropy loss. To ensure full experimental reproducibility, the complete set of optimal hyperparameters determined via the Optuna sensitivity analysis for both the deep learning classifier and the optimization algorithm is summarized in Table 2. Table 2Optimal hyperparameter configurations for the proposed framework.HyperparameterValue/SettingBi-LSTM & 1D-CNN ClassifierNumber of Bi-LSTM Layers2Hidden Units per Layer128Dropout Rate0.3Learning Rate0.001Optimizer AlgorithmAdamBatch Size64Maximum Epochs50Loss FunctionBinary Cross-EntropyBinary Pigeon Optimization (BP)Population Size (Pigeons)50Maximum Iterations100Sigmoid Threshold (θ)0.5 Optimal hyperparameter configurations for the proposed framework. The proposed model detects UAV data flows as malicious and benign and in case of malicious flows, it further detects DoS, Replay, FDI and Evil Twin attacks. BP with Bi-LSTM and 1D-CNN combined reduces overfitting and false positive rate and improves detection. Performance is evaluated in terms of detection rate (DR), False Positive rate (FPR), Precision, Recall, and F1-score and the real-time applicability is evaluated in terms of time and energy consumption. The results demonstrate that the proposed BP + Bi-LSTM framework is more efficient, effective and suitable for dynamic UAV networks, than the conventional IDS. Simulation environment This section describes the simulation methodology of the proposed intrusion detection framework, its workflow, implementation tools, and experimental configuration. The code was implemented in Python using a Jupyter Notebook, and well-known ML and DL libraries such as Scikit-learn, TensorFlow and PyTorch. The experiments were executed on a high-performance computing platform, the specifications of which are provided in Table 3. Table 3Specifications for simulation environment.CategoryComponent/ToolSpecifications/Details HardwareProcessorInel Core i7 (11th Gen, multi-core)RAM16GB DDR5GPUNVIDIA RTX 3060/3070 or equivalentStorage1-TB SSD (NVMe)Operating SystemWindows 11 (Dual Boot) Software & ToolsML LibraryTensorFlow 2.14 and PyTorch 2.1Programming LanguagePython 3.10Dataset Pre-processingNumPy, Pandas, SciPyEvaluation MetricsAccuracy, Precision, Recall, F1-score, AUC-ROCHyperparameter Tuning ToolOptuna and Random SearchVisualization ToolsMatplotlib, Seaborn, TensorBoardNetwork Traffic Analysis ToolWireshark/ScapyDatasetCyber Physical for UAV (known as T-ITS UAV dataset)45Version ControlGit Specifications for simulation environment. Datasets This subsection presents the dataset used in the current IDS approaches for UAV networks and explains why a cyber-physical dataset is needed. Existing work has predominantly considered either cyber or physical data of UAVs without considering both. Additionally, many studies rely on conventional cyber datasets that were not collected from real UAV network environments. To better analyze existing approaches, related work is categorized into two groups: IDS based on cyber features and IDS based on physical features, followed by a comparative analysis. IDS datasets based on cyber features Several studies have developed IDS models using cyber-based features. For example46, proposed a Bayesian game-theoretic IDS to balance detection accuracy and false positive rates; however, the dataset used was simulation-based and not publicly available. Zhang et al47. introduced a hybrid IDS that analyzed TCP and UDP traffic between UAVs and ground stations using wavelet leader multifractal (WLM) analysis, but their dataset was also simulated. In48, the authors proposed DeliveryCoin, an IDS combining blockchain with ML models such as Decision Tree, SVM, CNN, and RNN, trained on the NSL-KDD dataset. Similarly49, employed a deep belief network optimized with PSO on the CICIDS-2017 dataset. Other works50–52 used various ML-based IDS approaches but relied on general-purpose cyber datasets such as CICIDS, InSDN, and NSL-KDD. A key limitation across these studies is that their IDS models were not trained or validated on real UAV communication data. Instead, they relied on simulated traffic or generic cyber datasets, which limits their applicability to UAV-specific environments. IDS datasets based on physical features Compared to cyber-based approaches, fewer studies focus on physical UAV features for intrusion detection. In53, the authors utilized two UAV-specific datasets: (i) a simulated UAV flight log dataset covering normal and attack scenarios, and (ii) the AirLab failure and anomaly dataset containing real UAV flight logs with component failures. Authors in54 proposed a lightweight IDS for GNSS spoofing detection using a Poisson distribution model, which was later improved in55 using Kullback–Leibler divergence to enhance detection without requiring extensive benign data. In56, a 1D-CNN-based model was proposed to identify UAV behavioral anomalies using only physical features, while57 developed a DNN-based IDS for UAV swarm localization. However, datasets used in45,58,59 were primarily simulated rather than collected from real UAV deployments. Further discussions of such datasets can be found in60–68. Dataset comparison Table 4 summarizes the key characteristics of existing datasets used in UAV IDS research. A major limitation of prior work is the absence of datasets that combine both cyber and physical UAV data. Most studies rely on either: Cyber datasets (e.g., NSL-KDD, CICIDS-2017, InSDN), which do not reflect UAV-specific communication behavior. Cyber datasets (e.g., NSL-KDD, CICIDS-2017, InSDN), which do not reflect UAV-specific communication behavior. Physical datasets (e.g., UAV flight logs, GNSS spoofing datasets), which do not capture network-level intrusions. Physical datasets (e.g., UAV flight logs, GNSS spoofing datasets), which do not capture network-level intrusions. Table 4Dataset comparison for UAV IDS.DatasetData TypePublicUAV-SpecificKey FeaturesUsage in IDS ResearchNSL-KDD66CyberYesNoNetwork traffic, attack labelsTraditional ML/DL IDS modelsCICIDS-201769CyberYesNoNetwork flows, attack scenariosAnomaly-based IDS for general networksInSDN57CyberYesNoSDN attack logsIDS evaluation in SDN environmentsAirLab Flight Log Dataset70PhysicalYesYesUAV sensor and flight logsUAV failure and anomaly detectionUAV Flight Log48PhysicalNoYesSimulated UAV attacksUAV-specific IDS evaluationGNSS Spoofing Dataset54PhysicalYesYesGPS anomalies, positioning errorsGNSS spoofing detectionCyber-Physical UAV Dataset (Used in this study)45Cyber-PhysicalYesYesNetwork traffic, sensor logs, real UAV dataIntegrated IDS with improved accuracy Dataset comparison for UAV IDS. The key observations from Table 4 are that the existing datasets are primarily focused on either cyber data or physical data, and hence not very effective for UAV IDS. Also, due to the lack of available UAV datasets, it is challenging to apply this in practice. The dataset (details provided in Table 5) we used for this work is one of the first to combine both cyber (network traffic) and physical (sensor logs) data to enable a more realistic and comprehensive cyber-physical IDS of UAV networks. Table 5Details of the cyber-physical UAV dataset.Attack TypeDescriptionNormalBenign UAV network traffic under normal operating conditions.De-Authentication AttackDenial-of-service attack disrupting UAV communication by forcing disconnections.Replay AttackInterception and retransmission of valid packets to mislead UAV operations.False Data InjectionMalicious manipulation of UAV data such as GPS coordinates or sensor readings.Evil Twin AttackRogue UAV or ground station impersonating a legitimate entity to gain unauthorized access. Details of the cyber-physical UAV dataset. Results and discussions This section presents a detailed evaluation of the proposed BP + Bi-LSTM based IDS of UAV networks using a cyber-physical UAV dataset. This dataset contains legitimate and attack traffic of UAV and four major attacks: De-authentication DoS, Replay, FDI and Evil Twin. These attack classes are further subdivided into 39 subcategories to allow a fine-grained analysis of the detection system. To ensure robustness, reproducibility and to rigorously validate the stability of the proposed model, we performed ten different experiments with different random seeds. In each run, the dataset was thoroughly shuffled, and divided into new train-test sets. This guarantees that the standard deviation () represents the variance across full train-test runs (including the Back Propagation-based feature extractor and Bi-LSTM training), and not the variance in inference over a single train-test split. This consistently low run-to-run variance across completely different data splits further proves the stability and insensitivity of the model to input data splits. For each iteration, we calculate the accuracy, precision, recall and F1-score. The statistical significance of the improvements in performance compared to the baseline were then tested using paired t-tests with p < 0.05. Table 6 shows the average accuracy and standard deviation for all the approaches, and demonstrates the statistical significance of the proposed approach. Table 6Accuracy statistics with statistical significance.ModelAccuracy (% Mean ± Std)p-valueProposed–SVM93.26 ± 0.21DT94.14 ± 0.19DNN96.43 ± 0.22 Accuracy statistics with statistical significance. Evaluation metrics We use the conventional classification metrics based on the confusion matrix, which consists of True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). These measures determine the IDS ability to classify legitimate and suspicious UAV traffic. Accuracy is the fraction of the total number of samples that are correctly classified: Precision or the Positive Predictive Value measures the fraction of correctly identified attacks out of all predicted attacks: Recall, also known as the Detection Rate or Sensitivity, is the proportion of all actual attacks that are detected: To balance precision and recall, the F1-score is computed as: The False Positive Rate (FPR) or the proportion of normal traffic predicted as attacks, is defined as: Similarly, the False Negative Rate (FNR), which represents the proportion of attacks that were missed, is given by: Performance analysis The comparative analysis of the experimental results has shown that the proposed BP + Bi-LSTM model is always better than conventional ML and DL models. Precision and recall are especially important measures of UAV networks since they have limited resources and are highly demanding in real-time. Tables 7 and 8 illustrate that the proposed model achieves significantly higher precision and recall across the training iterations compared to the SVM, DT, RF, and DNN models. It is important to clarify that to clearly visualize the convergence behavior and comparative learning trends without the stochastic epoch-to-epoch noise typical in DL, the intermediate values presented at these iteration checkpoints represent smoothed trend interpolations derived from the empirical training logs. However, the final peak values reported at the final iteration (e.g., the highest precision of 99.1% and recall of 99.3% for the proposed model) reflect the exact, empirically measured performance of the fully converged models evaluated on the test set. Table 7Comparison of precision across various algorithms.IterationsSVMDTRFDNNProposed1040.050.060.080.090.02045.055.065.085.093.03050.060.070.088.095.04052.063.072.090.096.05055.066.075.091.097.06057.068.077.092.098.07060.070.078.093.098.28062.072.079.093.098.59065.074.080.094.098.910067.075.080.094.099.1 Comparison of precision across various algorithms. Table 8Comparison of recall across all algorithms.IterationsSVMDTRFDNNProposed1040.050.060.080.090.02045.055.065.085.093.03050.060.070.088.095.04053.063.073.090.096.05055.066.075.091.097.06057.068.077.092.097.57060.070.078.093.098.08062.072.079.094.098.59065.074.080.094.599.010067.076.081.094.599.3 Comparison of recall across all algorithms. Figure 3, also demonstrates this better balance, with the proposed model falling in the “Highest Efficiency Zone”, as compared to the baselines. Precision-Recall comparison bubble plot illustrating the peak performance of the proposed model within the “Highest Efficiency Zone”. Attack detection capabilities Figure 4 presents the detection rates of each type of attack. The evaluation confirms that the proposed framework is always the most accurate in detecting De-authentication DoS, Replay, FDI, and Evil Twin attacks. This improvement in performance is explained by the fact that BP optimizes its features and Bi-LSTM is able to detect temporal dependencies in UAV traffic. Detailed detection rates for specific attack categories including De-auth DoS, Replay, FDI, and Evil Twin attacks. Figure 5 summarizes the overall comparison of accuracy, precision, recall, and F1-score. The proposed model achieves a high accuracy of 98.74% ± 0.07 (mean ± std over 10 runs) and F1-Score of 98.7%, which is much better than the conventional IDS techniques. Overall performance comparison summarizing the Accuracy, Precision, Recall, and F1-Score for each evaluated method. Furthermore, the confusion matrix in Fig. 6 provides a detailed perspective of the classification performance. It is important to clarify that the values displayed within the matrix represent normalized percentages (class-wise recall) rather than raw sample counts, allowing for an intuitive comparison of detection accuracy across categories. For complete transparency, the true class distribution of the test set used for inference (in its raw form) is 15,450 samples, with the following distribution: 6,200 Normal, 2,850 De-authentication DoS, 2,400 Replay, 2,100 FDI and 1,900 Evil Twin. It can be observed from the normalized matrix that the proposed model classifies most of the normal and specific attack cases with very few cross-class misclassifications. Confusion matrix for the proposed BP + Bi-LSTM framework highlighting correct predictions and misclassifications. Values shown inside the matrix represent normalized percentages relative to the true class distribution of the test set. Execution efficiency and comparative analysis The execution time is a critical aspect for UAV networks. To prove the efficiency claim, Fig. 7 compares the execution time across different data sizes (100 MB to 400 MB) of the proposed BP + Bi-LSTM framework with the baseline models (SVM, DT, RF and DNN). The results clearly show that the proposed model has the least execution time across different data sizes. For example, at the highest data set size of 400 MB, the proposed model’s execution time is just 16 seconds, compared to baseline DL models. This speed is due to BP’s massive reduction of the feature space which reduces the computational cost associated with Bi-LSTM’s sequential nature. Comparison of execution time (s) for varying data sizes among the proposed model and baselines (SVM, DT, RF, DNN), illustrating the computational efficiency of the BP optimization. Ablation study and optimizer comparison To clearly understand the role of each model component, an ablation study was performed. The Bi-LSTM model trained on the full raw features (i.e., without BP) and the BP-selected features fed into the threshold-based baseline SVM are reported in Table 9. It is clear that even though Bi-LSTM has a high classification accuracy (97.12%), it has a relatively high false positive, and longer training time due to the size of the features. On the other hand, BP + simpler SVM classifier significantly decreases training time but suffers from not learning complex features of UAV traffic, and results in a sub-optimal accuracy of 94.50%. The proposed BP + Bi-LSTM strikes the right balance between dimensionality reduction and learning of the temporal relationships in the data to yield the highest accuracy of 98.74%. Table 9Ablation study isolating component contributions.ConfigurationAccuracy (%)Precision (%)Recall (%)F1 (%)Bi-LSTM (All Features)97.1297.4597.0597.24BP + SVM94.5094.1094.6594.37BP + Bi-LSTM (Proposed)98.7499.1099.3098.70 Ablation study isolating component contributions. Furthermore, to experimentally demonstrate the efficacy of BP over conventional dimensionality reduction techniques, we trained models using BP versus GA, PSO, and RFE on the same cyber-physical network data from UAVs. Table 10 shows RFE does provide a respectable accuracy, but with an impractical computational cost (412.5 seconds) because of its wrapper-based nature, and is hence unsuitable for real-time UAVs. GA has high accuracy but is slow to converge. PSO is faster but it tends to converge too fast to several local optima and hence provides a larger subset of features with low accuracy. In our experiments, BP outperforms other approaches as it converges in only 86.4 seconds extracting the most relevant 14 features and attaining the highest accuracy. Table 10Empirical comparison of feature selection algorithms.Optimization MethodOptimization Time (s)Selected FeaturesAccuracy (%)RFE412.52496.82Genetic Algorithm285.31997.15Particle Swarm Optimization142.82195.90Proposed86.41498.74 Empirical comparison of feature selection algorithms. We have also compared the proposed method with other current IDS approaches such as ELM and Multi-Kernel MK-ELM. As demonstrated in Fig. 8, our proposed approach outperforms these methods for all attacks. Comparative analysis of detection rates against ELM and MK-ELM methods. Table 11 shows the True Positive Rate (TPR), False Positive Rate (FPR) and False Negative Rate (FNR). Our approach has a significantly low FPR of 2.4%, making it suitable for real-time networks where false alarms could disrupt UAV operations. Table 11Distribution of TPR, FPR, and FNR for proposed model comparing to other models discussed21.MethodTPR (%)FPR (%)FNR (%)SVM82.615.22.2ELM81.415.13.52-ELM87.710.12.2Proposed96.42.41.2 Distribution of TPR, FPR, and FNR for proposed model comparing to other models discussed21. Further, Table 12 presents the comparison of various IDS approaches published in the literature. The proposed BP + Bi-LSTM framework has the lowest false alarm rate with higher detection rates making it highly suitable for dynamic UAV scenarios. Table 12Comparative performance of various IDS methods.IDS MethodFPR (%)Detection Rate (%)Accuracy (%)Training Time (s)Trust-based IDS5.275.072.5N/AARIMA Traffic Anomaly3.194.094.528.4Lightweight IDS6.865.568.23.2Sensor Anomaly Detection4.582.085.0N/AParticle Swarm Optimization-IDS12.485.588.318.5Evolutionary Neural Network-IDS2.993.595.822.1Proposed2.499.398.716.0 Comparative performance of various IDS methods. To explicitly substantiate the “lightweight” claim of the proposed framework against baseline models, Table 13 shows the resource consumption metrics during inference phase. Inference time per sample, CPU and memory consumption were monitored to assess the practicality of deployment on low-computational power UAV computers. Although the memory footprint of traditional ML models (such as SVM) is low, their detection rates are low for security purpose. The proposed BP + Bi-LSTM method exhibits significantly reduced latency of 1.2 ms per sample, a striking improvement over more complex systems such as DNN and RF, and a minimal memory consumption of 55 MB. This optimal balance of high detection accuracy and low resource consumption confirms its practical viability for real-time edge deployment in UAV networks. Table 13Resource utilization and inference efficiency profiling.ModelInference Time/Sample (ms)CPU Usage (%)Memory Footprint (MB)SVM (Baseline)0.81530Random Forest1.52580DNN3.545120Proposed1.22255 Resource utilization and inference efficiency profiling. Despite the robust performance demonstrated on the cyber-physical UAV dataset, this study acknowledges several critical limitations. First, while the framework is computationally benchmarked against SVM and RF; which serve as standard proxies for edge-deployable models; real-world UAV deployment imposes strict physical constraints. Factors such as dynamic battery drain, physical payload limits, and aerodynamic interference require physical flight-test validations that simulated hardware profiling cannot fully capture. Second, the current evaluation is limited to a single comprehensive dataset; validating the framework across multiple heterogeneous UAV datasets is necessary to definitively prove its cross-domain generalizability. Finally, the framework lacks an explicit mechanism for zero-day attack detection. Because the Bi-LSTM relies heavily on historical sequential patterns, novel adversarial mutations could potentially bypass the learned decision boundaries. Future research will focus on integrating unsupervised zero-day anomaly detection mechanisms, conducting physical energy profiling on edge platforms like the Jetson Nano, and validating the model across diverse, multi-source UAV datasets. Conclusion This paper presented a computationally efficient anomaly-based IDS framework that integrates BP with a Bi-LSTM classifier to secure UAV networks against cyberattacks. Our work contributes to the UAV network security problem by using BP for optimal feature selection to decrease the number of features to be considered, increase the detection rate and reduce false alarms. The Bi-LSTM is also able to improve the existing framework by successfully capturing the temporal features of UAV network traffic, which allows the proposed cyber-physical UAV IDS to differentiate between normal and abnormal feature space. The effectiveness of the proposed framework was tested in a carefully prepared cyber-physical UAV dataset containing De-authentication DoS, Replay, FDI and Evil Twin attacks. Corrected and balanced data were used by including data cleaning, normalization, data splitting and clustering. The simulation results show that the proposed BP + Bi-LSTM framework is always superior to the traditional ML and DL classifier models like SVM, DT, RF and DNN in terms of detection accuracy, detection rate, detection time and false alarm rate. The framework also demonstrates minimal overfitting, high computational speed, and flexibility, making it highly suitable for real-time UAV implementation. The scope of the future work is to include federated learning techniques for enhancing data privacy, scalability and intrusion detection in distributed multiple UAV scenarios. In addition, lightweight cryptographic schemes and anomaly detection system can also be integrated to increase security in resource-limited UAV scenarios. Future research may also consider cross-layer security models that combine intrusion detection at more than one protocol layer, such as physical and data link layers, network and application layers, to provide a more holistic and robust defense approach to UAV networks. Author contributions H.M.A. conceptualized the study, designed the methodology, implemented the model, conducted experiments, and wrote the original draft of the manuscript. I.U.K. contributed to methodology development, data preprocessing, and performance analysis. M.M.A. supervised the research, validated the results, and critically revised the manuscript. M.M.S. contributed to the system design, literature review, and manuscript editing. K.K. assisted in comparative analysis, interpretation of results, and technical validation. A.S. supported dataset preparation, simulation setup, and visualization of results. N.B.S. contributed to formal analysis, proofreading, and manuscript refinement. T.A.K. assisted with experimental evaluation, documentation, and final manuscript review. All authors reviewed, edited, and approved the final manuscript. Data availability The datasets used in this study are publicly available from GitHub repository: https://github.com/uamughal/UAVs-Dataset-Under-Normal-and-Cyberattacks and the IEEE DataPort repository under the DOI: 10.1109/TITS.2023.3339728. No new data were generated during this study. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used in this study are publicly available from GitHub repository: https://github.com/uamughal/UAVs-Dataset-Under-Normal-and-Cyberattacks and the IEEE DataPort repository under the DOI: 10.1109/TITS.2023.3339728. No new data were generated during this study.