A novel deep learning approach for intrusion detection in maritime radar networks Authors: Md Mostofa Nurannabi Shakil, Md. Alamgir Hossain, Md. Samiul Islam, Nidadavolu Venkat Durga Sai Siva Vara Prasad Raju Published: 2025 ABSTRACT In recent years, maritime radar networks have become essential for ensuring the safety and security of maritime operations. However, with the increased interconnectivity of these systems, they have also become vulnerable to cyber-attacks, posing significant risks to critical infrastructure. Traditional intrusion detection systems (IDS) often struggle to detect sophisticated and evolving attacks in real-time due to their reliance on manual feature extraction and shallow machine learning techniques. This research addresses this gap by introducing MARINERNet, a deep learning-based intrusion detection system designed specifically for maritime radar networks. The proposed system uses a novel architecture that integrates 1D convolutional layers, squeeze-and-excitation blocks, and residual connections to automatically extract relevant features from raw radar network data, enhancing detection accuracy without manual intervention. MARINERNet is evaluated on both binary and multiclass classification tasks, demonstrating state-of-the-art performance. Specifically, the model achieves 98.52% accuracy for multiclass classification and perfect accuracy for anomaly detection (binary classification). The approach is scalable, capable of handling large datasets, and adaptable to real-time intrusion detection, making it suitable for deployment in dynamic radar environments. This research not only provides an effective solution for detecting intrusions in maritime radar networks but also contributes to the broader field of cybersecurity by offering a robust, deep learning-based approach that can be applied to other network systems. Introduction This section provides an overview of the growing cybersecurity challenges in maritime radar networks and introduces the motivation, scope, and contributions of this research. It also contextualizes the study within existing intrusion detection systems and highlights the novel aspects of MARINERNet. Background and motivation Maritime radar signal processing involves several critical operations that convert raw electromagnetic reflections into meaningful navigational insights. Radar systems emit radio frequency pulses and then process the reflected signals to detect objects, determine their range, velocity, and direction1. A fundamental technique in radar processing is Doppler shift analysis, which helps differentiate between moving and stationary targets by measuring frequency changes in the returned signals. This is crucial for identifying other vessels, buoys, or obstacles in real-time2. Another important process is clutter suppression, which mitigates unwanted echoes caused by sea waves, rain, or atmospheric interference. Techniques such as Constant False Alarm Rate (CFAR) and adaptive filtering are widely employed to enhance signal clarity. Additionally, pulse compression techniques, using matched filtering, allow radar systems to achieve high resolution without increasing transmitted power, improving the detection of small or distant targets3,4. These signal processing techniques form the foundation of modern maritime radar systems and are essential for reliable navigation and situational awareness, especially in dynamic marine environments where both safety and security depend on accurate and timely radar interpretation5. Maritime intrusion refers to any unauthorized access, manipulation, or disruption of maritime navigation systems, particularly radar and communication networks, which compromises the safety, security, and operational integrity of maritime operations6,7. With the increasing interconnectivity of modern maritime radar networks, vessels and port facilities are becoming more reliant on real-time radar and sensor data to ensure safe navigation, collision avoidance, and monitoring of surrounding environments. While these networks provide critical situational awareness, their growing complexity and connectivity have also introduced significant cybersecurity vulnerabilities, making them susceptible to sophisticated cyberattacks that can manipulate radar data, disrupt operations, or mask the presence of vessels8,9. The motivation for this research stems from the recognition that traditional intrusion detection systems (IDS) and conventional machine learning approaches are often inadequate for modern maritime radar networks10,11. These approaches typically rely on manual feature engineering and shallow models, which struggle to detect novel or evolving attack strategies in real-time. Additionally, the operational constraints of maritime environments such as high traffic volume, dynamic network conditions, and imbalanced data distributions, demand intrusion detection systems that are not only highly accurate but also scalable and capable of real-time deployment12,13. Recent studies have highlighted the increasing frequency and sophistication of maritime cyberattacks, including Denial-of-Service (DoS) attacks that disrupt radar communications, object manipulation attacks that add, remove, or relocate vessels on radar displays, and transformation attacks such as scaling, rotation, and translation of radar signals14,15. These threats underscore the urgent need for advanced detection mechanisms capable of automatically learning discriminative features from raw radar data without extensive manual intervention. Maritime radar systems play a crucial role in ensuring the safety and security of vessels at sea by providing real-time situational awareness. These systems operate by emitting electromagnetic waves (radio waves) and analyzing the reflected signals from objects within their range, such as other vessels, landmasses, and environmental obstacles. The radar signals are processed to detect and classify targets based on their distance, speed, and direction. In maritime radar signal processing, raw radar returns are typically filtered, transformed, and analyzed using various techniques to ensure effective target detection and reduce interference from the environment6,16. One key technique in maritime radar signal processing is Doppler processing, which allows for the measurement of the velocity of objects relative to the radar7. By analyzing the Doppler shifts in the reflected signals, the radar system can distinguish moving targets from stationary ones, which is particularly important for detecting moving vessels or obstacles in the vicinity of the ship. This processing is crucial for collision avoidance and navigation in congested waters. Additionally, clutter suppression is vital to improving radar performance. Radar systems are often affected by environmental noise such as sea waves, rain, or other non-target reflections (clutter). Advanced techniques like adaptive filtering and CFAR (Constant False Alarm Rate) algorithms are employed to reduce this clutter, ensuring that only relevant targets, such as other ships or obstacles, are detected and tracked4,17. Another essential aspect of maritime radar signal processing is target detection and tracking. After noise and clutter have been filtered out, radar systems use detection algorithms such as matched filtering and moving target indication (MTI) to identify potential targets18. These algorithms assess the strength, position, and movement of detected objects. Following detection, tracking filters like Kalman filters or particle filters are employed to predict the future trajectory of targets based on their current movement. These tracking filters are essential for continuous monitoring and avoiding false positives in dynamic maritime environments. In addition, signal modulation techniques like pulse compression are used to enhance the radar’s resolution without compromising its detection range, allowing for better identification of smaller or distant objects19,20. However, the integration of maritime radar systems into broader communication networks has increased their vulnerability to cyber threats. As radar systems become more connected, the risk of cyberattacks that could manipulate radar data or disrupt operations grows. For example, attackers exploit vulnerabilities in the communication protocols, causing disruption in radar operations or falsifying radar data. The need for robust and adaptive intrusion detection systems (IDS). Challenges in maritime radar intrusion detection Intrusion detection in maritime radar networks faces several key challenges. First, the dynamic and heterogeneous nature of radar traffic, including real-time signals and control messages across multiple protocols, creates complex and noisy data that is difficult to analyze accurately. Second, there is a class imbalance, as benign traffic dominates while attack instances such as DoS, object manipulation, and transformation attacks are rare, making detection of critical threats challenging. Third, traditional IDS and shallow machine learning approaches rely heavily on manual feature engineering, which is time-consuming and often fails to capture evolving or sophisticated attacks. Finally, ensuring real-time detection and scalability across diverse radar scenarios adds further complexity21–23. These challenges necessitate an adaptive, automated, and high-performance solution capable of handling high-dimensional, imbalanced, and dynamic radar data, which is the primary motivation behind MARINERNet. Objectives of the research and major contributions The primary objective of this research is to develop MARINERNet, a deep learning-based intrusion detection system for maritime radar networks. The system aims to enhance detection accuracy by automatically identifying both known and novel cyberattacks, such as DoS, scaling, and object manipulation, while minimizing false positives and negatives. It automates feature extraction using 1D convolutional layers and squeeze-and-excitation blocks, ensuring real-time performance and scalability for large datasets. Additionally, MARINERNet supports both binary and multiclass classification tasks, making it adaptable for various operational requirements and providing a robust solution for securing maritime infrastructure against cyber threats. This research makes several significant contributions to the field of maritime radar network security and intrusion detection. Major contributions are listed below: We introduced MARINERNet, a deep learning-based intrusion detection system that integrates 1D convolutions, squeeze-and-excitation blocks, and residual connections to effectively detect intrusions in maritime radar networks.We integrated deep convolutional layers for automatic feature extraction from raw radar network data, improving the system’s ability to detect novel attacks without manual feature engineering.The designed approach is scalable, capable of handling large radar network datasets, and adaptable for real-time intrusion detection in various radar systems.The proposed approach is effective for both binary and multiclass classification, making it adaptable for a wide range of intrusion detection tasks. We introduced MARINERNet, a deep learning-based intrusion detection system that integrates 1D convolutions, squeeze-and-excitation blocks, and residual connections to effectively detect intrusions in maritime radar networks. We introduced MARINERNet, a deep learning-based intrusion detection system that integrates 1D convolutions, squeeze-and-excitation blocks, and residual connections to effectively detect intrusions in maritime radar networks. We integrated deep convolutional layers for automatic feature extraction from raw radar network data, improving the system’s ability to detect novel attacks without manual feature engineering. We integrated deep convolutional layers for automatic feature extraction from raw radar network data, improving the system’s ability to detect novel attacks without manual feature engineering. The designed approach is scalable, capable of handling large radar network datasets, and adaptable for real-time intrusion detection in various radar systems. The designed approach is scalable, capable of handling large radar network datasets, and adaptable for real-time intrusion detection in various radar systems. The proposed approach is effective for both binary and multiclass classification, making it adaptable for a wide range of intrusion detection tasks. The proposed approach is effective for both binary and multiclass classification, making it adaptable for a wide range of intrusion detection tasks. Organization of the paper The remainder of this paper is structured as follows: Sect. 2 reviews related work on maritime radar intrusion detection and deep learning-based IDS approaches. Section 3 presents the proposed methodology, including dataset preparation, preprocessing, and the MARINERNet architecture. Section 4 details the experimental setup, performance metrics, and results for both multiclass and binary classification, followed by a discussion of insights. Finally, Sect. 5 concludes the study, highlighting the significance of the findings and potential directions for future research. Related works The security of maritime radar networks has become a critical area of research due to the increasing reliance on digital radar systems for navigation, collision avoidance, and situational awareness. Traditional approaches to intrusion detection in these networks largely relied on signature-based methods and rule-based systems, which are inherently limited in their ability to detect novel or sophisticated cyberattacks. For example, conventional Intrusion Detection Systems (IDS) often monitor for known attack signatures such as Denial-of-Service (DoS) or simple data manipulations, but fail when attackers deploy novel transformations like scaling, rotation, or translation of radar objects. These limitations have motivated researchers to explore machine learning (ML) and deep learning (DL) techniques, which can automatically learn patterns from raw network traffic and identify anomalies without predefined rules. Several studies have demonstrated the effectiveness of shallow machine learning methods for radar network security. Techniques such as Random Forests, Support Vector Machines (SVMs), and k-Nearest Neighbors (k-NN) have been applied to network traffic features extracted from radar systems, achieving moderate success in detecting DoS attacks and simple anomalies. However, these methods typically require manual feature engineering, which is both labor-intensive and domain-specific. Moreover, they struggle with high-dimensional radar data and imbalanced datasets, common in realistic maritime environments where benign traffic dominates and malicious activity is sparse19,24,25. Hossain et al.23 proposed a semi-supervised learning method for marine radar security using a Self-Training Classifier with Random Forest, achieving 99.99% accuracy on the RadarPWN dataset. While effective in controlled settings, the method has key limitations: it overfit to the dataset, lacks testing under diverse maritime conditions, and faces challenges in real-time deployment and integration with existing radar systems. To overcome these limitations, researchers have increasingly turned to deep learning architectures, particularly Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), which can automatically extract hierarchical features from sequential or spatial data. For instance, CNNs have been effectively applied to classify radar signals and detect adversarial manipulations in network traffic without manual intervention. Similarly, RNNs and Long Short-Term Memory (LSTM) networks have shown promise in capturing temporal dependencies in radar traffic, improving the detection of stealthy or time-dependent attacks26–28. These studies underscore the potential of deep learning to address both feature engineering challenges and complex attack patterns in maritime radar networks. In addition to deep learning models, recent works have explored attention mechanisms and residual learning to enhance detection performance. Attention mechanisms allow models to focus on the most informative features, improving detection accuracy for subtle attacks, while residual connections help train deeper networks without suffering from vanishing gradients. Integrating these techniques has been shown to significantly improve both multiclass and binary classification performance in intrusion detection tasks29,30. However, most existing approaches have focused either on small-scale datasets or specific attack types, limiting their applicability to realistic maritime radar environments, which include a mixture of benign traffic and diverse attack scenarios. Despite these advances, gaps remain in developing scalable, real-time intrusion detection systems capable of handling high-dimensional radar data while maintaining high detection accuracy across multiple attack types. Current models often lack adaptability to new or evolving attacks, and few provide both binary and multiclass detection capabilities in a single framework. This research addresses these gaps by introducing MARINERNet, a deep learning-based IDS that combines 1D convolutions, squeeze-and-excitation blocks, and residual connections. MARINERNet is designed to automatically extract features from raw radar traffic, handle large datasets efficiently, and perform robust detection across both binary and multiclass classification tasks, making it suitable for deployment in operational maritime radar networks. Methodology This section outlines the detailed methodology employed in this research to develop and evaluate the MARINERNet intrusion detection system for maritime radar networks. The methodology is divided into several stages, including data preprocessing, model development, training, and evaluation. The complete development pipeline of the proposed approach is shown in Fig. 1. Fig. 1Development Pipeline of the Complete Approach. Development Pipeline of the Complete Approach. Dataset Preparation The dataset used in this study is the RadarPWN dataset31, which is a specialized collection designed to evaluate the resilience of maritime radar systems against various cyberattack techniques. The dataset is derived from simulated maritime environments, including scenarios such as Rostock, Simple Estuary, and Santa Catalina. These environments represent different radar operating conditions captured during simulated maritime operations using tools like Bridge Command, OpenCPN, and the Radar Attack Toolkit (RAT). To provide an intuitive understanding of the input data used for training and evaluation, we include visual examples of each attack class from the RadarPWN dataset. As shown in Fig. 2, the dataset includes seven distinct attack types: Denial of Service, Scaling, Rotation, Translation, Object Addition, Object Removal, and Object Relocation. Each attack alters the radar display in a specific way, simulating realistic and potentially hazardous adversarial scenarios. For instance, DoS attacks (Fig. 2b) obscure the radar signal using random or blank noise, while scaling (Fig. 2c) and rotation (Fig. 2d) affect the perceived distance and orientation of targets. More advanced attacks such as object addition, removal, or relocation (Fig. 2f–h) manipulate surrounding objects in a context-aware manner, mimicking deceptive tactics against navigation systems. Table 1 summarizes the variations of each attack type and their inclusion in the dataset. These visuals and metadata provide insight into the complexity and realism of the radar intrusion scenarios that MARINERNet is designed to detect. Fig. 2Visual examples of radar signal manipulation for each of the seven attack classes in the RadarPWN dataset31. Visual examples of radar signal manipulation for each of the seven attack classes in the RadarPWN dataset31. These are critical tools for simulating maritime radar systems and evaluating their security under cyberattack conditions. These applications are widely used in maritime navigation and situational awareness, providing essential functionalities that enhance the operational efficiency and safety of vessels at sea. Bridge Command is a maritime navigation simulation tool that mimics the operational environment of a vessel’s bridge, supporting radar and other navigation aids. It is used for simulating various radar-based scenarios in controlled environments to assess vulnerabilities in radar communication protocols32,33.OpenCPN (Open Chart Plotter Navigator)34,35 is an open-source navigation software that provides real-time chart plotting, radar data integration, and situational awareness for maritime operations. It is often employed in combination with other navigation systems to ensure safe and efficient maritime travel.Radar Attack Toolkit (RAT)36 is a set of tools designed specifically for simulating cyberattacks on radar systems. It allows researchers to model various types of attacks, such as Denial-of-Service (DoS), object manipulation, and radar signal transformation, which are critical for testing the resilience of radar systems in the face of adversarial threats. Bridge Command is a maritime navigation simulation tool that mimics the operational environment of a vessel’s bridge, supporting radar and other navigation aids. It is used for simulating various radar-based scenarios in controlled environments to assess vulnerabilities in radar communication protocols32,33. Bridge Command is a maritime navigation simulation tool that mimics the operational environment of a vessel’s bridge, supporting radar and other navigation aids. It is used for simulating various radar-based scenarios in controlled environments to assess vulnerabilities in radar communication protocols32,33. OpenCPN (Open Chart Plotter Navigator)34,35 is an open-source navigation software that provides real-time chart plotting, radar data integration, and situational awareness for maritime operations. It is often employed in combination with other navigation systems to ensure safe and efficient maritime travel. OpenCPN (Open Chart Plotter Navigator)34,35 is an open-source navigation software that provides real-time chart plotting, radar data integration, and situational awareness for maritime operations. It is often employed in combination with other navigation systems to ensure safe and efficient maritime travel. Radar Attack Toolkit (RAT)36 is a set of tools designed specifically for simulating cyberattacks on radar systems. It allows researchers to model various types of attacks, such as Denial-of-Service (DoS), object manipulation, and radar signal transformation, which are critical for testing the resilience of radar systems in the face of adversarial threats. Radar Attack Toolkit (RAT)36 is a set of tools designed specifically for simulating cyberattacks on radar systems. It allows researchers to model various types of attacks, such as Denial-of-Service (DoS), object manipulation, and radar signal transformation, which are critical for testing the resilience of radar systems in the face of adversarial threats. Table 1Attack types and their variations included in the RadarPWN dataset31.Attack TypeExample (Fig. 2)VariationsTotal Samples (Σ)Denial of Service (DoS)Figure 2(b)blank, random, freeze, turn off4ScalingFigure 2(c)pixel-wise, protocol header2RotationFigure 2(d)pixel-wise, protocol header, context-aware3TranslationFigure 2(e)pixel-wise, context-aware2Object AdditionFigure 2(f)context-aware (incl. AIS & ARPA)1Object RemovalFigure 2(g)context-aware (incl. AIS & ARPA)1Object RelocationFigure 2(h)context-aware (incl. AIS & ARPA)1 Table 1Attack types and their variations included in the RadarPWN dataset31.Attack TypeExample (Fig. 2)VariationsTotal Samples (Σ)Denial of Service (DoS)Figure 2(b)blank, random, freeze, turn off4ScalingFigure 2(c)pixel-wise, protocol header2RotationFigure 2(d)pixel-wise, protocol header, context-aware3TranslationFigure 2(e)pixel-wise, context-aware2Object AdditionFigure 2(f)context-aware (incl. AIS & ARPA)1Object RemovalFigure 2(g)context-aware (incl. AIS & ARPA)1Object RelocationFigure 2(h)context-aware (incl. AIS & ARPA)1 Attack types and their variations included in the RadarPWN dataset31. The dataset31 captures both benign traffic (normal radar operations) and malicious activity in the form of various cyberattacks on radar communication protocols. The primary attacks simulated include: Denial-of-Service (DoS): Aiming to disrupt radar operations by overwhelming the system with excessive requests, preventing the radar from functioning correctly. The DoS class in the dataset is the least frequent, making up only 8.77% of the total data.Scaling: Modifies the size of the radar image or objects detected by the radar, making it appear larger or smaller than they actually are. This attack comprises 11.88% of the total dataset.Addition: Involves adding false objects to the radar image, such as ships or obstacles that don’t exist, creating confusion. The Addition class also contributes 11.86% of the data.Rotation: Alters the orientation of detected objects in the radar display, potentially misleading radar operators. It accounts for 11.73% of the dataset.Move: This attack changes the position of objects detected by the radar, making them appear to move inappropriately. It forms 11.64% of the data.Translation: Involves shifting the radar image, affecting the positioning of the objects. Translation makes up 11.57% of the dataset.Removal: Involves the removal of objects from the radar display, making it appear as though certain objects are no longer present. The Removal attack class represents 10.25% of the dataset. Denial-of-Service (DoS): Aiming to disrupt radar operations by overwhelming the system with excessive requests, preventing the radar from functioning correctly. The DoS class in the dataset is the least frequent, making up only 8.77% of the total data. Denial-of-Service (DoS): Aiming to disrupt radar operations by overwhelming the system with excessive requests, preventing the radar from functioning correctly. The DoS class in the dataset is the least frequent, making up only 8.77% of the total data. Scaling: Modifies the size of the radar image or objects detected by the radar, making it appear larger or smaller than they actually are. This attack comprises 11.88% of the total dataset. Scaling: Modifies the size of the radar image or objects detected by the radar, making it appear larger or smaller than they actually are. This attack comprises 11.88% of the total dataset. Addition: Involves adding false objects to the radar image, such as ships or obstacles that don’t exist, creating confusion. The Addition class also contributes 11.86% of the data. Addition: Involves adding false objects to the radar image, such as ships or obstacles that don’t exist, creating confusion. The Addition class also contributes 11.86% of the data. Rotation: Alters the orientation of detected objects in the radar display, potentially misleading radar operators. It accounts for 11.73% of the dataset. Rotation: Alters the orientation of detected objects in the radar display, potentially misleading radar operators. It accounts for 11.73% of the dataset. Move: This attack changes the position of objects detected by the radar, making them appear to move inappropriately. It forms 11.64% of the data. Move: This attack changes the position of objects detected by the radar, making them appear to move inappropriately. It forms 11.64% of the data. Translation: Involves shifting the radar image, affecting the positioning of the objects. Translation makes up 11.57% of the dataset. Translation: Involves shifting the radar image, affecting the positioning of the objects. Translation makes up 11.57% of the dataset. Removal: Involves the removal of objects from the radar display, making it appear as though certain objects are no longer present. The Removal attack class represents 10.25% of the dataset. Removal: Involves the removal of objects from the radar display, making it appear as though certain objects are no longer present. The Removal attack class represents 10.25% of the dataset. The Benign class, representing normal radar operations, dominates the dataset with 40,251 samples (22.31%) of the total data. While the dataset exhibits a natural class imbalance, with benign traffic being more prevalent than attack traffic, the diverse attack types help to create a well-rounded, high-fidelity dataset that mirrors real-world conditions where attacks are less frequent but still significant. Each attack type is characterized by various network features such as packet length, timestamp, protocol type, and inter-arrival times, which were extracted using the Scapy library and processed into 66 features per sample. These features are designed to provide rich, discriminative information about both benign and malicious network traffic. The dataset’s imbalanced nature reflects the typical scenario in maritime radar operations, where malicious activity is sparse compared to normal operations. The amount of data of different attacks with their percentage after processing is displayed in Fig. 3. Data preprocessing Data preprocessing is a critical step in machine learning and deep learning tasks as it directly influences the performance of the model. In this research, preprocessing was essential for ensuring the quality of the dataset before feeding it into the MARINERNet intrusion detection model. This section describes the processes employed to handle missing data, encode categorical features, scale numerical features, and split the data into training and test sets. Fig. 3Class distribution of the dataset after the processing, showing the number of samples for each class. Class distribution of the dataset after the processing, showing the number of samples for each class. Handling missing data Handling missing data is crucial for ensuring that the model receives clean and complete input. In this research, missing data in both numerical and categorical columns was addressed through imputation techniques tailored to their respective data types. For numerical columns, missing values were replaced with the median value of the respective column. The median was chosen over the mean as it is more robust to outliers and ensures a more stable and representative imputation, especially in datasets with skewed distributions. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X=\{{x}_{1},{x}_{2},\dots\:,{x}_{n}\}$$\end{document} represent a numerical feature with missing values. The median for this feature is calculated as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:M=\left\{\begin{array}{c}middle\:value\:of\:sorted\left(X\right),\:\:\:if\:n\:is\:odd\\\:\frac{middle\:two\:values\:of\:sorted\left(X\right)}{2}\:,\:\:if\:n\:is\:even\end{array}\right.$$\end{document} For categorical columns, missing values were imputed with the mode (most frequent value) of the column. The mode was chosen since categorical data usually follows a non-normal distribution, and the mode provides the most representative value for imputation. For a categorical feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:C=\{{c}_{1},{c}_{2},\dots\:,{c}_{n}\}$$\end{document}, the mode is computed as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m={arg}max\left(\sum\:_{i=1}^{n}{1}_{\left\{{c}_{i}=x\right\}}\:\right)$$\end{document} Where, 1 is an indicator function that returns 1 if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{i}=x$$\end{document} and 0 otherwise. After applying the imputation strategies, missing values were handled effectively, allowing the dataset to be fully utilized for training without any loss of information. Feature encoding and scaling Feature encoding and scaling are essential preprocessing steps, especially when the dataset contains both categorical and numerical data. These steps ensure that the features are in an appropriate format for the deep learning model and that all input features are scaled consistently to improve convergence during training. For categorical features, Label Encoding was applied. Label encoding transforms each category into a unique integer. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:C=\{{c}_{1},{c}_{2},\dots\:,{c}_{n}\}$$\end{document} be a categorical feature with distinct values. The LabelEncoder assigns each category a unique integer from 0 to −1, such that:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:LabelEncoder\left({c}_{i}\right)=i\:for\:i\in\:[0,m-1]$$\end{document} For numerical features, scaling is crucial as it ensures that each feature contributes equally to the model’s performance, especially when using gradient-based optimization. In this research, StandardScaler was applied to standardize the numerical features. Standardization rescales the data to have a mean of 0 and a standard deviation of 1. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X=\{{x}_{1},{x}_{2},\dots\:,{x}_{n}\}$$\end{document} represent a numerical feature. The standardized value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{std}$$\end{document} is computed as:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{std}=\frac{{x}_{i}-\:\mu\:}{\sigma\:}$$\end{document} Where is the mean and is the standard deviation of the feature. This standardization step prevents features with larger numerical ranges from dominating the learning process, ensuring that all features contribute equally. Data splitting Data splitting is the process of partitioning the dataset into training, validation, and test sets. This step is essential to ensure that the model can be properly trained, validated, and evaluated without overfitting or bias. In this research, the dataset was split into training and test sets. The training set was used to train the MARINERNet model, while the test set was used to evaluate its performance on unseen data. To ensure that the class distribution was maintained across both sets, stratified sampling was employed. Stratified sampling ensures that the proportions of each class in the original dataset are preserved in the training and test sets. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D=\left\{\right({x}_{1},{y}_{1}),({x}_{2},{y}_{2}),\dots\:,({x}_{n},{y}_{n}\left)\right\}$$\end{document} represent the full dataset, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}$$\end{document} is the feature vector and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}$$\end{document} is the corresponding class label. The data is split into training \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({D}_{train}\right)$$\end{document} and test \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({D}_{test}\right)$$\end{document} sets as follows: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{train},{D}_{test}=StratifiedSplit(D,{test}_{size}=0.2)$$\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}$$\:{test}_{size}=0.2$$\end{document} indicates that 20% of the data is reserved for testing and the rest for training. The stratified split ensures that each class has the same proportion in both the training and test sets, which is particularly important for imbalanced datasets. The data was then further split into training and validation sets during the model training process. The validation set (20% of the training data) was used to tune the model and monitor its performance during training to prevent overfitting. Model architecture MARINERNet is a deep learning-based intrusion detection system designed to automatically learn relevant features from raw radar network traffic data. Unlike traditional intrusion detection systems (IDS) that rely on manual feature engineering, MARINERNet utilizes a deep convolutional neural network (CNN) to perform feature extraction and classification directly from the raw input data. The model leverages a 1D convolutional layer for extracting sequential patterns from the data, squeeze-and-excitation (SE) blocks for recalibrating channel-wise feature responses, and residual connections to improve gradient flow and enable deeper architectures. List of hyperparameters with their values for the development of the proposed approach are presented in. Table 2, and the architecture in Fig. 4. We proposed MARINERNet, a deep learning-based intrusion detection system that leverages a Convolutional Neural Network (CNN) architecture to automatically extract relevant features from raw radar network data. The model consists of three 1D convolutional layers, each designed to capture hierarchical features from the radar traffic data. These layers, combined with squeeze-and-excitation blocks and residual connections, enhance the network’s ability to detect intrusions without requiring manual feature engineering. The architecture is as follows: i.Input Layer: The input to the model is a 2D array of shape (,), where is the number of samples (data points), and is the number of features (attributes) for each sample. The data is reshaped to have dimensions (,, 1) to be compatible with the 1D convolution operation.ii.Conv1D Layers: The model uses convolutional layers to capture spatial patterns across the data, applying filters to extract relevant features.iii.Squeeze-and-Excitation Block: A module that enhances feature selection by learning attention weights for each channel, allowing the network to focus on more important features.iv.Residual Connections: These connections allow the model to learn deeper representations by bypassing certain layers, facilitating gradient flow and mitigating the vanishing gradient problem.v.Fully Connected Layer: After the convolutional layers, a global average pooling layer is used to aggregate feature maps, which is then passed through a fully connected layer for final classification.vi.Output Layer: The model outputs a probability distribution over the classes using the softmax activation function. i.Input Layer: The input to the model is a 2D array of shape (,), where is the number of samples (data points), and is the number of features (attributes) for each sample. The data is reshaped to have dimensions (,, 1) to be compatible with the 1D convolution operation. Input Layer: The input to the model is a 2D array of shape (,), where is the number of samples (data points), and is the number of features (attributes) for each sample. The data is reshaped to have dimensions (,, 1) to be compatible with the 1D convolution operation. ii.Conv1D Layers: The model uses convolutional layers to capture spatial patterns across the data, applying filters to extract relevant features. Conv1D Layers: The model uses convolutional layers to capture spatial patterns across the data, applying filters to extract relevant features. iii.Squeeze-and-Excitation Block: A module that enhances feature selection by learning attention weights for each channel, allowing the network to focus on more important features. Squeeze-and-Excitation Block: A module that enhances feature selection by learning attention weights for each channel, allowing the network to focus on more important features. iv.Residual Connections: These connections allow the model to learn deeper representations by bypassing certain layers, facilitating gradient flow and mitigating the vanishing gradient problem. Residual Connections: These connections allow the model to learn deeper representations by bypassing certain layers, facilitating gradient flow and mitigating the vanishing gradient problem. v.Fully Connected Layer: After the convolutional layers, a global average pooling layer is used to aggregate feature maps, which is then passed through a fully connected layer for final classification. Fully Connected Layer: After the convolutional layers, a global average pooling layer is used to aggregate feature maps, which is then passed through a fully connected layer for final classification. vi.Output Layer: The model outputs a probability distribution over the classes using the softmax activation function. Output Layer: The model outputs a probability distribution over the classes using the softmax activation function. Fig. 4Layer-by_Layer Architecture of the Proposed MARINERNet. Layer-by_Layer Architecture of the Proposed MARINERNet. Conv1D layers and feature extraction Convolutional layers are pivotal in MARINERNet for automatically extracting hierarchical features from the radar network traffic data. Specifically, 1D convolutional layers (Conv1D) are used, which are well-suited for sequential data such as time-series or radar signal data. The Conv1D operation slides a kernel/filter across the data, computing the dot product between the filter and a local region of the input, effectively capturing local dependencies. Given an input sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X=\{{x}_{1},{x}_{2},\dots\:,{x}_{F}\}$$\end{document}, here is the number of features, and a convolutional filter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:W=\{{w}_{1},{w}_{2},\dots\:,{w}_{k}\}$$\end{document}, the convolution operation is performed as follows:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}=\:\sum\:_{j=1}^{k}{w}_{j}{x}_{i+j-1}+b$$\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}$$\:{y}_{i}$$\end{document} is the output of the convolution at position, is the size of the convolutional kernel, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{j}$$\end{document} is the weight (filter) applied at position, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i+j-1}$$\end{document} is the input feature at position \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i+j-1$$\end{document}, is a bias term added to the convolution. Through this operation, the model captures local patterns across the features, such as the temporal or sequential relationships in the radar data. For MARINERNet, Conv1D layers are applied in successive stages: i.The first Conv1D layer applies a small kernel to capture basic features from the raw data.ii.Further Conv1D layers increase the complexity by applying larger kernels, enabling the model to extract more abstract and higher-level features as the depth of the network increases. i.The first Conv1D layer applies a small kernel to capture basic features from the raw data. The first Conv1D layer applies a small kernel to capture basic features from the raw data. ii.Further Conv1D layers increase the complexity by applying larger kernels, enabling the model to extract more abstract and higher-level features as the depth of the network increases. Further Conv1D layers increase the complexity by applying larger kernels, enabling the model to extract more abstract and higher-level features as the depth of the network increases. The number of filters in each convolutional layer increases progressively (e.g., 32, 64, etc.) to capture increasingly complex patterns in the data. Squeeze-and-excitation block The squeeze-and-excitation (SE) block is a key component introduced to enhance the representational power of the network by adaptively recalibrating the feature map channels. It works by focusing on the most relevant features in the feature map, suppressing the less informative ones. This is achieved through two main steps: squeeze and excitation. i. Squeeze: Global average pooling is applied to the output of the convolutional layers to summarize the feature map into a channel descriptor. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X\in\:{R}^{H\times\:W\times\:C}$$\end{document} be the output from the convolutional layers, where H is the height (number of samples), is the width (number of features), and is the number of channels (feature maps). The squeeze operation computes the channel-wise global average:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{c}=\frac{1}{H\:\times\:W}\sum\:_{i=1}^{H}\sum\:_{j=1}^{W}{X}_{i,\:j,\:c}$$\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}$$\:{s}_{c}$$\end{document} is the average of channel across all spatial locations. ii. Excitation: The channel descriptors are passed through two fully connected layers to learn a set of attention weights that indicate the importance of each channel. The excitation step applies the learned weights to recalibrate the original feature map:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{c}={X}_{c}\cdot\:\sigma\:({W}_{2}\cdot\:ReLU({W}_{1}\cdot\:{s}_{c}\left)\right)$$\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}$$\:{X}_{c}$$\end{document} is the feature map of channel, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} is the sigmoid activation function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{1},{W}_{2}$$\end{document} are the weights of the fully connected layers. This recalibration allows the network to focus on more informative channels and suppress less useful ones, improving the detection capability of the model. Residual connections Residual connections, introduced in ResNet, are used to facilitate the training of deep networks by mitigating the vanishing gradient problem. These connections allow the network to learn the residual mapping instead of the direct mapping, which simplifies the learning process. In MARINERNet, residual connections are applied after certain convolutional layers to enhance gradient flow and improve convergence. Given an input and an output () from a convolutional block, the residual connection creates an identity shortcut:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Y=F\left(X\right)+X$$\end{document} Where, F(X) is the output of the convolutional block, is the input to the block, and is the output after adding the identity shortcut. This simple addition ensures that the network can learn the residual mapping \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F\left(X\right)$$\end{document} instead of the full transformation, which helps in training deeper networks. By adding the shortcut connection, the model is encouraged to focus on learning the difference between the input and output, rather than the entire transformation, making it easier to optimize the weights. The residual connection helps the MARINERNet model to remain efficient and avoid degradation in performance as the number of layers increases, especially when dealing with complex radar network data. The MARINERNet architecture is a sophisticated design that leverages powerful techniques such as 1D convolutional layers, squeeze-and-excitation blocks, and residual connections to build an effective and efficient intrusion detection system. These components work together to automatically learn hierarchical features, focus on the most relevant features through attention mechanisms, and enable the training of deeper models without encountering optimization issues. This combination of techniques ensures that MARINERNet delivers state-of-the-art performance in detecting intrusions in maritime radar networks. Table 2MARINERNet architecture summary for both binary and multiclass classification tasks.Layer NameTypeOutput Shape# ParamsConnectionsInputInput Layer(None, 66, 1)0–Conv1D-11D Convolution(None, 66, 32)128InputBatchNorm-1Batch Normalization(None, 66, 32)128Conv1D-1Conv1D-21D Convolution(None, 66, 64)6,208BatchNorm-1GlobalAvgPool-1Global Avg Pooling(None, 64)0Conv1D-2Dense-SE-1Dense (SE Block)(None, 8)520GlobalAvgPool-1Dense-SE-2Dense (SE Block)(None, 64)576Dense-SE-1SE MultiplyMultiply(None, 66, 64)0Conv1D-2, Dense-SE-2Conv1D-31D Convolution(None, 66, 64)12,352SE MultiplyDropout-1Dropout(None, 66, 64)0Conv1D-3Residual AddAdd (Residual Link)(None, 66, 64)0Dropout-1, SE MultiplyGlobalAvgPool-2Global Avg Pooling(None, 64)0Residual AddDense-1Fully Connected(None, 32)2,080GlobalAvgPool-2Dropout-2Dropout(None, 32)0Dense-1Output LayerDense (Softmax)(None, 8/2)264Dropout-2 Table 2MARINERNet architecture summary for both binary and multiclass classification tasks.Layer NameTypeOutput Shape# ParamsConnectionsInputInput Layer(None, 66, 1)0–Conv1D-11D Convolution(None, 66, 32)128InputBatchNorm-1Batch Normalization(None, 66, 32)128Conv1D-1Conv1D-21D Convolution(None, 66, 64)6,208BatchNorm-1GlobalAvgPool-1Global Avg Pooling(None, 64)0Conv1D-2Dense-SE-1Dense (SE Block)(None, 8)520GlobalAvgPool-1Dense-SE-2Dense (SE Block)(None, 64)576Dense-SE-1SE MultiplyMultiply(None, 66, 64)0Conv1D-2, Dense-SE-2Conv1D-31D Convolution(None, 66, 64)12,352SE MultiplyDropout-1Dropout(None, 66, 64)0Conv1D-3Residual AddAdd (Residual Link)(None, 66, 64)0Dropout-1, SE MultiplyGlobalAvgPool-2Global Avg Pooling(None, 64)0Residual AddDense-1Fully Connected(None, 32)2,080GlobalAvgPool-2Dropout-2Dropout(None, 32)0Dense-1Output LayerDense (Softmax)(None, 8/2)264Dropout-2 MARINERNet architecture summary for both binary and multiclass classification tasks. Model training and optimization Model training and optimization are critical stages in developing any deep learning model, as they ensure that the model not only learns from the data but also generalizes well to unseen instances. In this research, the training process for the MARINERNet model involves the use of modern optimization techniques and hyperparameter tuning to maximize its performance for the task of intrusion detection in maritime radar networks. This section provides a detailed description of the training process, hyperparameter tuning, loss function and metrics, and model evaluation strategy. Training process The training process for MARINERNet involves the iterative optimization of the model’s parameters (weights and biases) based on the input data. The dataset is divided into training and validation sets, with the training set used to update the model parameters, and the validation set used to monitor the model’s performance during training. The goal is to minimize the loss function while improving the model’s accuracy in classifying both benign and attack traffic. Data Preparation The training data is preprocessed, including scaling of numerical features and encoding of categorical variables as described in Sect. 3.2. This ensures that all input data is in a suitable form for the neural network. The training and validation sets are split in a stratified manner, ensuring that the class distribution is preserved across both sets. Model training The MARINERNet model is trained using backpropagation and the Nadam optimizer. Backpropagation involves calculating the gradients of the loss function with respect to each model parameter and adjusting the parameters to reduce the error between predicted and true labels. The optimizer uses an adaptive learning rate to adjust the model’s parameters efficiently. Specifically, Nadam is a combination of Adam and Nesterov momentum and helps the model converge faster. The general training update rule is. 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}={\theta\:}_{t}-\eta\:\cdot\:{\nabla\:}_{\theta\:}L\left({\theta\:}_{t}\right)$$\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}$$\:{\theta\:}_{t}$$\end{document} represents the model parameters at time step t, η is the learning rate, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\nabla\:}_{\theta\:}L\left({\theta\:}_{t}\right)$$\end{document} is the gradient of the loss function with respect to the parameters at time step, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L\left({\theta\:}_{t}\right)$$\end{document} is the loss function. Batch size and epochs A batch size of 32 samples was chosen for training, allowing the model to update the weights frequently and speed up convergence. The model was trained for up to 200 epochs to ensure sufficient training. Early stopping is employed to monitor the model’s validation loss and stop the training once the validation performance stops improving, preventing overfitting. Callbacks The EarlyStopping callback halts the training if the validation loss does not improve for 10 consecutive epochs. The ReduceLROnPlateau callback is used to decrease the learning rate by a factor of 0.5 if the validation loss stagnates. Hyperparameter tuning Hyperparameter tuning is essential for optimizing the model’s architecture and training process. It involves searching for the best combination of hyperparameters that lead to the highest model performance. Key hyperparameters tuned in this study include: Learning rate The learning rate controls the step size taken during the optimization process. A higher learning rate might lead to faster convergence but could overshoot the minimum, while a lower learning rate ensures gradual progress but take more epochs to converge. Number of Filters and Kernel Size: In the convolutional layers, the number of filters and the kernel size were fine-tuned. The number of filters determines how many features maps the network will learn at each convolutional layer. The kernel size controls the size of the local region the filter will examine, affecting how much local context is considered when extracting features. Dropout rate Dropout is a regularization technique used to prevent overfitting. The dropout rate specifies the proportion of neurons to be randomly deactivated during training. Experimenting with different dropout rates allows the model to generalize better to unseen data. Batch size The batch size determines how many samples are processed before the model’s internal parameters are updated. In this study, a batch size of 32 was used, but this was also fine-tuned during the optimization process. The tuning was performed through grid search or random search over predefined parameter spaces. Once the best hyperparameters were identified, the model was retrained using those optimal settings. Loss functions The choice of the loss function and evaluation metrics is crucial to guide the model toward effective learning and accurate predictions. In this research, the following loss function and metrics were used: Loss function Since the problem is both binary and multi-class classification problem, categorical crossentropy was used as the loss function. This loss function computes the difference between the predicted probability distribution and the true probability distribution (ground truth) for each class. The categorical crossentropy loss is calculated as:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L=-\:\sum\:_{c=1}^{C}{y}_{c}{log}\left({\widehat{y}}_{c}\right)$$\end{document} Where, C is the number of classes, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{c}$$\end{document} is the true label for class (0 or 1), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{c}$$\end{document} is the predicted probability for class. The training and optimization of the MARINERNet model were designed with a focus on achieving high performance and generalization. Through an iterative training process, careful hyperparameter tuning, and the use of appropriate loss functions and evaluation metrics, the model was able to learn meaningful features and accurately detect intrusions in maritime radar networks. The proposed MARINERNet is formally summarized in Algorithm 1, which outlines the complete workflow for intrusion detection in maritime radar networks. The algorithm takes preprocessed radar network flows as input, applies a sequence of 1D convolutional layers, a squeeze-and-excitation (SE) attention mechanism, and residual temporal blocks for robust feature extraction, followed by global average pooling and fully connected layers for classification. The network is trained using categorical cross-entropy loss with the Nadam optimizer, while early stopping and learning rate reduction strategies ensure efficient convergence. Algorithm 1 highlights the iterative forward and backward passes over mini-batches, weight updates, and real-time prediction steps, providing a comprehensive representation of the proposed approach. Experimental results and analysis This section presents the experimental setup, evaluation metrics, and a comprehensive analysis of MARINERNet’s performance on maritime radar network intrusion detection. The results are analyzed both for multiclass and binary classification scenarios, highlighting the robustness, accuracy, and practical applicability of the proposed approach. Experimental setup The experimental setup for this research was implemented using Google Colab, which provides a cloud-based environment with access to free hardware resources. The hardware accelerator used was the CPU, which was suitable for the scale of this deep learning task. The primary libraries utilized include TensorFlow 2.7 and Keras for building and training the deep learning model, while Scikit-learn was used for preprocessing tasks such as feature scaling with StandardScaler, label encoding with LabelEncoder, and data splitting using train_test_split. NumPy was employed for numerical operations and data manipulation, and Pandas was used for reading and cleaning the dataset. Matplotlib and Seaborn were utilized for visualizing training progress and evaluating model performance through metrics like confusion matrices. Performance metrics To evaluate the effectiveness of MARINERNet, we employed standard classification metrics: Accuracy: Measures overall correctness of predictions.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$\end{document} Where, TP = True Positives (correctly predicted attack instances), TN = True Negatives (correctly predicted benign instances), FP = False Positives (incorrectly predicted benign instances as attack, FN = False Negatives (incorrectly predicted attack instances as benign). Precision: Proportion of true positives among predicted positives.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Precision=\frac{TP}{TP+FP}$$\end{document} Recall (Sensitivity): Proportion of true positives among actual positives.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Recall=\frac{TP}{TP+FN}$$\end{document} F1 Score: Harmonic mean of precision and recall.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F1=2*\frac{Precision\times\:Recall}{Precision+Recall}$$\end{document} AUC (Area Under the ROC Curve): Indicates the model’s ability to distinguish between classes. The ROC curve plots true positive rate (TPR) vs. false positive rate (FPR). We also use confusion matrices to visualize class-wise performance and identify misclassification trends. These performance metrics were crucial in assessing the effectiveness of MARINERNet for detecting intrusions in maritime radar networks. Results for multiclass classification The results for the multiclass classification task demonstrate the impressive performance of MARINERNet in detecting various types of intrusions and benign activities within the maritime radar network dataset. At Epoch 120, the model achieved a training accuracy of 96.54% and a validation accuracy of 98.50%, indicating strong generalization capabilities. The loss function also showed significant improvement, with the training loss at 0.0861 and validation loss at 0.0568. The learning rate was adjusted during training, settling at 3.1250e-05 by this epoch, contributing to improved model convergence. These early-stage results highlight the model’s effectiveness in capturing and distinguishing between various network activity types, making it well-suited for real-time intrusion detection in maritime radar networks. The evaluation metrics further confirm the model’s robustness. The overall accuracy for the multiclass classification task was 98.52%, with a precision of 98.57%, recall of 98.52%, and F1 score of 98.52%. These results indicate that the model not only correctly identifies intrusions and benign traffic but also does so with high consistency, as demonstrated by the near-perfect F1 score. The model achieves an excellent balance between precision and recall, making it effective for both detecting attacks and minimizing false positives. The classification report presented in Table 3 provides a detailed breakdown of the model’s performance across different classes. For example, the model performs exceptionally well on the Addition and DoS classes, achieving 99.98% precision for Addition and 99.94% precision for DoS, along with high recall values of 98.32% and 98.61%, respectively. These results indicate that the model effectively identifies even the most subtle attack types. However, the Scaling class shows slightly lower performance, with a precision of 99.15% and recall of 94.96%, which suggests a potential area for further model improvement, particularly in handling this class. Overall, the macro average F1 score of 98.52% and the weighted average F1 score of 98.52% emphasize the model’s ability to perform well across a diverse range of classes, making it a powerful tool for detecting intrusions in maritime radar networks. Table 3Classification report for multiclass Classification.ClassPrecisionRecallF1-ScoreAddition0.99980.98320.9914Benign0.97820.99990.9889DoS0.99940.98610.9927Move0.99900.98880.9939Removal0.99570.99080.9932Rotation0.99660.98060.9886Scaling0.99150.94960.9701Translation0.93580.99070.9624 Accuracy 0.9852 Macro Avg 0.98700.98370.9852 Weighted Avg 0.98570.98520.9852 Table 3Classification report for multiclass Classification.ClassPrecisionRecallF1-ScoreAddition0.99980.98320.9914Benign0.97820.99990.9889DoS0.99940.98610.9927Move0.99900.98880.9939Removal0.99570.99080.9932Rotation0.99660.98060.9886Scaling0.99150.94960.9701Translation0.93580.99070.9624 Accuracy 0.9852 Macro Avg 0.98700.98370.9852 Weighted Avg 0.98570.98520.9852 Classification report for multiclass Classification. Figure 5 shows the training accuracy and training loss curves for the MARINERNet model during the training process for multiclass classification. The plot on the left demonstrates how the model’s accuracy improves over the epochs for both the training and validation sets. Initially, the training accuracy increases rapidly, stabilizing at a high value as the model continues to learn. Similarly, the plot on the right illustrates the training and validation loss, where the loss drops sharply during the initial epochs, indicating that the model is quickly learning to minimize the error. Afterward, both curves continue to converge, signaling the model’s effective learning and generalization over the course of training. The small gap between training and validation accuracy, as well as the minimal difference in loss values, reflects the model’s good generalization capabilities and robust performance on unseen data. These curves provide an overall view of the model’s convergence behavior and its ability to effectively distinguish between the different classes. Fig. 5Training curves showing accuracy (left) and loss (right) for both the training and validation sets. Training curves showing accuracy (left) and loss (right) for both the training and validation sets. Figure 6 presents the confusion matrix for the MARINERNet model’s performance in the multiclass classification task. The matrix shows the distribution of true versus predicted labels across all classes. Each row corresponds to the true class, and each column represents the predicted class. The values along the diagonal (from top-left to bottom-right) indicate the number of correctly classified instances for each class, showing high accuracy in classifying most of the categories. For instance, the Benign class (8049 instances) and the DoS class (3120 instances) are correctly identified with minimal misclassification. However, some off-diagonal values indicate misclassifications, such as Addition (48 misclassified instances) and Translation (7 misclassified instances), though these misclassifications are relatively low in comparison to the number of correct predictions. This suggests that the model performs well in differentiating between most classes, but there is room for improvement in recognizing certain types of attacks or activities, especially for the Scaling and Translation classes. The color intensity of the cells corresponds to the frequency of predictions, with darker shades indicating a higher number of instances. The matrix provides a comprehensive visualization of the model’s performance, demonstrating its ability to accurately identify the various classes in the dataset while revealing areas for potential refinement. Fig. 6Confusion matrix showing the performance of MARINERNet in multiclass classification. Confusion matrix showing the performance of MARINERNet in multiclass classification. The higher prediction error rates observed in the Benign and Translation classes, as shown in Table 3; Fig. 6, can be attributed to the imbalanced nature of the dataset, which reflects real-world maritime radar scenarios. The Benign class has a significantly larger number of instances compared to the attack classes, leading to potential false positives where normal traffic is misclassified as an attack. This class imbalance is inherent in operational environments, where benign traffic overwhelmingly dominates radar data. Similarly, Translation attacks, due to their subtle nature, are difficult to distinguish from benign fluctuations, further contributing to the errors. Binary classification results In the binary classification task, where the labels were simplified into Benign and Attack categories, MARINERNet demonstrated exceptional performance. The dataset was modified such that all non-Benign labels were reclassified as Attack, resulting in a total of 140,194 instances of Attack and 40,251 instances of Benign. This transformation made it easier to evaluate the model’s ability to detect benign activities versus intrusions. During Epoch 98 of training, the model achieved an impressive accuracy of 99.99% on the training set, with a validation accuracy of 100%. The loss values were exceptionally low, with the training loss at 3.6415e-04 and validation loss at 1.6314e-07, indicating that the model successfully minimized errors during the training process. The learning rate was adjusted to 1.0000e-05, helping the model converge to an optimal solution efficiently. The evaluation metrics in Table 4 for the binary classification were perfect, with accuracy, precision, recall, and F1 score all reaching a value of 1.0000, highlighting the model’s flawless ability to distinguish between benign and attack instances. The classification report further confirms the model’s exceptional performance, with both Attack and Benign classes achieving perfect precision, recall, and F1 scores. Figure 7 presents the training accuracy and loss curves for the binary classification task, where the model distinguishes between Benign and Attack labels. The accuracy curve on the left shows a rapid increase in both training and validation accuracy, with both curves approaching 1.0 around Epoch 20, indicating the model’s quick and efficient learning. The loss curve on the right demonstrates a significant drop in both training and validation loss in the initial epochs, followed by stabilization at very low values, highlighting the model’s ability to minimize errors effectively during training. The small gap between the training and validation curves further indicates that the model generalizes well to unseen data, without overfitting, making it highly reliable for real-time binary classification tasks in intrusion detection. Table 4Classification results for anomaly Detection.ClassPrecisionRecallF1-ScoreAttack1.00001.00001.0000Benign1.00001.00001.0000 Accuracy 1.0000 Macro Avg 1.00001.00001.0000 Weighted Avg 1.00001.00001.0000 Table 4Classification results for anomaly Detection.ClassPrecisionRecallF1-ScoreAttack1.00001.00001.0000Benign1.00001.00001.0000 Accuracy 1.0000 Macro Avg 1.00001.00001.0000 Weighted Avg 1.00001.00001.0000 Classification results for anomaly Detection. Fig. 7raining curves for binary classification showing accuracy (left) and loss (right) for both the training and validation sets. raining curves for binary classification showing accuracy (left) and loss (right) for both the training and validation sets. Figure 8 presents the confusion matrix for the binary classification task, where the labels are classified into Attack and Benign. The matrix shows that MARINERNet achieved perfect performance with no misclassifications. The Attack class was correctly predicted 28,039 times, and the Benign class was accurately predicted 8,050 times. No false positives or false negatives were observed, confirming the model’s exceptional ability to distinguish between benign activities and attacks in the radar network. The color intensity in the matrix highlights the magnitude of correct predictions, with darker shades indicating higher counts. Fig. 8Confusion matrix for binary classification, showing perfect classification with no misclassifications between Attack and Benign labels. Confusion matrix for binary classification, showing perfect classification with no misclassifications between Attack and Benign labels. Although MARINERNet achieved perfect accuracy (100%) in the binary classification (anomaly) task, several factors within the model design and training process help ensure that this result is not a symptom of overfitting but rather reflects genuine generalization. First, the model incorporates regularization techniques, including dropout layers and batch normalization, which prevent the network from memorizing training examples and instead encourage it to learn generalizable patterns. Dropout randomly deactivates neurons during training, reducing interdependency among features, while batch normalization stabilizes the learning process by normalizing intermediate activations. Second, early stopping and learning rate reduction on plateau callbacks were employed. Early stopping monitors the validation loss and halts training once no improvement is observed for a fixed number of epochs, thus preventing the model from continuing to train beyond its generalization point. The adaptive learning rate adjustment further refines convergence by lowering the step size when improvement stagnates. Third, stratified data splitting ensured that both the training and test sets contained representative distributions of benign and attack classes, maintaining data diversity and minimizing data leakage. The model’s validation accuracy consistently mirrored training accuracy during the process, confirming that the observed perfect performance extended to unseen samples rather than being restricted to the training data. Finally, the relatively compact model architecture and 1D convolutional structure contribute to a lower risk of overfitting by limiting model complexity while maintaining high expressive power through residual and squeeze-and-excitation connections. These design decisions collectively enhance MARINERNet’s robustness and ensure that its exceptional binary accuracy reflects a well-generalized model rather than overfitting. Table 5 presents a detailed comparison of the computational performance and model parameters for both the anomaly (binary) and multiclass classification tasks in MARINERNet. The table outlines the total, trainable, non-trainable, and optimizer parameters, as well as the training time, model size, and hardware configuration used during experimentation. For the anomaly (binary) classification task, MARINERNet consists of 66,049 total parameters, of which 21,994 are trainable, 64 are non-trainable, and 43,991 belong to the optimizer. For the multiclass task, the model comprises 66,643 total parameters, with 22,192 trainable, 64 non-trainable, and 44,387 optimizer parameters. The minor increase in total parameters in the multiclass model arises from the expanded output layer required to accommodate multiple attack classes. From a computational perspective, the training time was approximately 4,763 s for the anomaly task and 5,606 s for the multiclass task on a CPU-only (x86_64) setup with 12.67 GB RAM. Both configurations produced a compact model of only 0.32 MB, highlighting MARINERNet’s lightweight and resource-efficient architecture. These results demonstrate that the proposed framework maintains computational efficiency while achieving high accuracy, making it feasible for real-time deployment in maritime radar environments where hardware resources be constrained. Table 5Comparison of model parameters for anomaly and multiclass classification tasks in MARINERNet.MetricsAnomalyMulticlassTotal params66,049 (258.01 KB)66,643 (260.33 KB)Trainable params21,994 (85.91 KB)22,192 (86.69 KB)Non-trainable params64 (256.00 B)64 (256.00 B)Optimizer params43,991 (171.84 KB)44,387 (173.39 KB)Training Time (s)4763.145606.10Model Size (MB)0.320.32Hardware UsedCPU onlyCPU Infox86_64RAM Available12.67 GB Table 5Comparison of model parameters for anomaly and multiclass classification tasks in MARINERNet.MetricsAnomalyMulticlassTotal params66,049 (258.01 KB)66,643 (260.33 KB)Trainable params21,994 (85.91 KB)22,192 (86.69 KB)Non-trainable params64 (256.00 B)64 (256.00 B)Optimizer params43,991 (171.84 KB)44,387 (173.39 KB)Training Time (s)4763.145606.10Model Size (MB)0.320.32Hardware UsedCPU onlyCPU Infox86_64RAM Available12.67 GB Comparison of model parameters for anomaly and multiclass classification tasks in MARINERNet. Interpretability with explainable AI (XAI) While high predictive performance is crucial, model interpretability is equally essential for operational trust, particularly in safety-critical maritime environments. Deep learning models like MARINERNet can behave as “black boxes,” making it difficult for domain experts to understand the reasoning behind their predictions. To address this, we integrated Explainable AI (XAI) techniques to provide both local and global insights into the model’s decision process. We employ LIME (Local Interpretable Model-Agnostic Explanations) to visualize per-sample feature attributions and Permutation Feature Importance (PFI) to assess overall feature influence on prediction performance. These complementary approaches illuminate how MARINERNet distinguishes between benign radar operations and diverse intrusion scenarios, thereby enhancing transparency, accountability, and user confidence in real-world deployments. Figure 9 demonstrates how MARINERNet interprets a radar flow labeled as Attack with complete confidence (P = 1.00). The left column reports class probabilities; the central chart decomposes the contribution of each feature toward the “Attack” and “Benign” outcomes; the right panel lists normalized feature values. Key discriminative features, Total Backward Packets, Backward Packet Length Max, and Total Length of Backward Packets, push the decision decisively toward the Attack class, while milder influences from Backward IAT Min/Max provide limited benign support. Such explanations enhance operator trust by revealing the internal rationale behind high-confidence anomaly predictions. Fig. 9LIME-based local explanation for an anomaly detection instance (binary case). LIME-based local explanation for an anomaly detection instance (binary case). The bar plot in Fig. 10 highlights how each feature contributes either positively (green) or negatively (red) to the Benign hypothesis. Strong negative weights from Total Backward Packets and Backward Packet Length Max dominate, while moderate positive signals from Backward Packet Length Std/Mean and Inter-Arrival Time Std insufficiently offset them. Consequently, the flow is identified as malicious. This fine-grained view enables cybersecurity analysts to trace specific packet-level properties responsible for each classification outcome. Fig. 10Local explanation for the “Benign” class in binary anomaly detection. Local explanation for the “Benign” class in binary anomaly detection. Figure 11 summarizes the average ΔF1 degradation per feature when permuted, capturing MARINERNet’s global sensitivity profile for the binary setting. Consistent with the multiclass analysis, backward-direction metrics, packet counts, packet-length statistics, and inter-arrival time measures, emerge as the most influential. The alignment of feature importance across both binary and multiclass tasks confirms the model’s stability and interpretive coherence, validating that its predictions stem from physically grounded radar-traffic characteristics rather than coincidental correlations. Fig. 11Global permutation feature importance for binary anomaly detection. Global permutation feature importance for binary anomaly detection. Figure 12 illustrates MARINERNet’s decision process for a radar flow predicted as a “Move” intrusion event from multiclass classification. The left panel displays the model’s predicted class probabilities, confirming the model’s certainty (P = 1.00) in classifying the sample as Move. The middle bar chart contrasts features pushing the decision toward the “Benign” or “Non-Benign” direction, while the right panel lists the corresponding feature values. The most influential attributes, such as Backward Inter-Arrival Time (Bwd IAT), Backward Packet Length Mean/Standard Deviation, and Backward Segment Size Average, exert strong positive contributions toward the Move class. This interpretation verifies that MARINERNet distinguishes manipulation-based radar attacks by learning temporal and volumetric irregularities in backscattered radar communication. Fig. 12LIME-based local explanation for a multiclass intrusion instance. LIME-based local explanation for a multiclass intrusion instance. Figure 13 presents the top ten local feature contributions explaining why the selected radar instance was not classified as Benign. Positive (green) bars represent evidence supporting benign behavior, whereas negative (red) bars indicate factors discouraging the benign hypothesis. The contextual feature Scenario > 0.81 strongly suppresses the Benign probability, while low values of Bwd IAT Mean, Bwd Pkt Len Std, and Bwd Pkt Len Mean marginally support benignness. The predominance of negative evidence aligns with the final non-Benign (attack) prediction, illustrating the interpretability advantage of LIME for instance-level transparency. Fig. 13Local feature attribution for the “Benign” class (multiclass LIME analysis). Local feature attribution for the “Benign” class (multiclass LIME analysis). Figure 14 shows the reduction in weighted F1-score (ΔF1) after random shuffling of each input feature, quantifying its global contribution to model performance. Only a small subset of features causes substantial performance drops, confirming that MARINERNet relies primarily on semantically meaningful descriptors such as packet-length distributions, inter-arrival timing, and flow-level packet counts. The concentration of high-impact features at specific indices reflects the model’s ability to extract compact yet robust representations, underscoring its interpretability and computational efficiency. Fig. 14Global permutation feature importance for multiclass classification. Global permutation feature importance for multiclass classification. Comparison of the proposed approach with existing approaches Table 6 presents a comparative analysis between the proposed MARINERNet model and state-of-the-art maritime intrusion detection systems, MRS-PFIDS and FED-GEM-CN, both of which were evaluated on the same RadarPWN dataset. Hossain et al.23 also proposed a semi-supervised learning method using Random Forests for maritime radar intrusion detection. While this method achieved high performance, it was limited by its reliance on manual feature extraction and its inability to scale effectively with increasing radar data complexity. Hossain et al.21 also introduced MRS-PFIDS, a federated learning-driven semi-supervised intrusion detection framework that achieved an accuracy of 97.97%, precision of 95.99%, recall of 97.97%, and an F1-score of 96.97%. Although effective, its semi-supervised nature and reliance on feature-specific training limited its adaptability to diverse radar environments. Subsequently, Hossain11 proposed FED-GEM-CN, a dual-CNN-based federated model that incorporated contrastive cross-attention to enhance detection across distributed maritime systems, achieving 99.04% accuracy and an F1-score of 96.82%. In contrast, the proposed MARINERNet model achieved 98.52% accuracy, 98.70% precision, 98.37% recall, and an F1-score of 98.52%, demonstrating balanced and superior generalization without the need for federated aggregation or manual feature engineering. While the accuracy is slightly below the FED-GEM-CN benchmark, MARINERNet’s advantage lies in its lightweight, single-site deployability, automated feature extraction, and robust multiclass classification capability, which make it highly suitable for real-time intrusion detection in operational maritime radar networks. Table 6Comparison of marinernet’s performance with existing models using the same RadarPWN dataset.ModelAccuracyPrecisionRecallF1-scoreMRS-PFIDS [24]0.97970.95990.97970.9697FED-GEM-CN [11]0.99040.96890.96800.9682MARINERNet(Proposed)98.520.98700.98370.9852 Table 6Comparison of marinernet’s performance with existing models using the same RadarPWN dataset.ModelAccuracyPrecisionRecallF1-scoreMRS-PFIDS [24]0.97970.95990.97970.9697FED-GEM-CN [11]0.99040.96890.96800.9682MARINERNet(Proposed)98.520.98700.98370.9852 Comparison of marinernet’s performance with existing models using the same RadarPWN dataset. MARINERNet (Proposed) Discussion of results and insights The results obtained from the MARINERNet model’s performance in both multiclass and binary classification tasks demonstrate its remarkable efficacy in detecting intrusions and benign activities in maritime radar networks. The model achieved outstanding accuracy, with both multiclass and binary classification tasks reaching high precision, recall, and F1-scores. These metrics indicate that MARINERNet not only classifies intrusions accurately but also minimizes both false positives and false negatives, making it highly reliable for real-time intrusion detection. In the multiclass classification task, MARINERNet displayed an impressive accuracy of 98.52%, with near-perfect performance across most of the attack types, including Benign, DoS, and Addition. The confusion matrix highlighted the model’s strong ability to differentiate between various types of intrusions, with only a small number of misclassifications. However, certain attack types, like Scaling and Translation, showed slightly lower performance, with recall and precision scores dipping marginally compared to other classes. This suggests that while the model is highly effective, there is potential for improving its detection capabilities, particularly in recognizing more subtle attack types. When the task was simplified to a binary classification (Benign vs. Attack), the model’s performance was flawless, achieving perfect precision, recall, and F1-scores of 1.0000 for both classes. The confusion matrix confirmed that the model correctly classified all instances without any false positives or false negatives, demonstrating its robustness and reliability when distinguishing between benign traffic and attack activities. The ROC curves for the binary classification task also highlighted MARINERNet’s strong discriminatory power, with AUC scores nearing 1.0 for both classes, further confirming the model’s effectiveness in detecting attacks with minimal error. The training process for MARINERNet showed efficient convergence, with both training and validation accuracy quickly reaching near-perfect values and training loss dropping significantly in the early epochs. This rapid convergence suggests that the model is able to learn the underlying patterns in the data efficiently, even for complex, high-dimensional radar network traffic. One notable insight from the results is the model’s scalability and its ability to handle large datasets. The parameter analysis indicated that the number of trainable parameters remained manageable despite the model’s complexity. The model’s efficiency in training and performance on large datasets highlights its suitability for deployment in real-time systems where both speed and accuracy are critical. In terms of model optimization, the use of squeeze-and-excitation blocks and residual connections has contributed significantly to enhancing feature selection and improving gradient flow, which likely played a key role in the model’s ability to achieve high performance with relatively fewer parameters compared to other deep learning approaches. Additionally, the Nadam optimizer played an important role in speeding up convergence and minimizing the loss function effectively during training. The time complexity of the proposed MARINERNet model is primarily dominated by the computations within the Conv1D layers, particularly the dilated convolution with 64 filters, which process each input flow of radar features sequentially. While the squeeze-and-excitation (SE) blocks and residual connections enhance feature representation and gradient flow, they contribute minimally to the overall computational cost. During training, the forward and backward propagation through these convolutional layers scales approximately as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(N\cdot\:F\cdot\:K\cdot\:{C}_{in}\cdot\:{C}_{out})$$\end{document} per batch, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N$$\end{document} is the number of samples, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F$$\end{document} is the sequence length (number of features), KKK is the kernel size, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{in}$$\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}$$\:{C}_{out}$$\end{document}​ are the input and output channels, respectively. The dense layers and pooling operations add negligible overhead in comparison. Consequently, the overall training time scales linearly with the number of samples and epochs, while inference time per sample remains efficient, enabling real-time deployment in maritime radar networks. The compact architecture of MARINERNet contributes to its scalability. With approximately 66,000 total parameters (about 260 KB), the model size is significantly smaller compared to typical deep learning-based intrusion detection systems that have millions of parameters. This small footprint ensures that MARINERNet can be deployed on systems with limited computational resources, such as CPU-only embedded devices in maritime radar systems. Despite its modest size, the model efficiently processes large datasets, supporting high-volume radar traffic scenarios without exceeding memory or processing limits. The computational complexity of the model scales linearly with the number of samples, features, kernel sizes, and input/output channels. The core convolutional operations, combined with residual blocks and squeeze-and-excitation layers, have a linear relationship with input size. This ensures that as the number of radar data flows or features increases, the required computational resources grow proportionally. This scalability enables the model to be adapted for real-time intrusion detection in high-traffic environments, such as maritime radar networks, without compromising its performance. Empirical evidence further supports the scalability claim. When tested on a standard CPU accelerator in Google Colab, the binary classification model reached convergence in approximately 4,763 s, while the multiclass version took 5,606 s with around 180,000 samples for training and validation. This demonstrates the model’s ability to train efficiently even with large datasets. Inference time per sample remains under a few milliseconds, allowing the model to handle real-time traffic in operational environments. Moreover, MARINERNet is adaptable to increasing data volumes and new classes. The modular architecture of the proposed model, including 1D convolutional layers, squeeze-and-excitation blocks, and residual connections, allows for easy adjustments when adding new attack types or operational conditions. For instance, adding new classes only requires modifications to the final dense layer, with a negligible increase in the number of parameters (about 600 extra parameters). This demonstrates functional scalability, where new attack types can be incorporated without requiring a full re-architecture or significant computational overhead. The architecture of the model is designed with deployment in mind, particularly in resource-constrained maritime radar environments. Given the constraints of radar systems, such as limited memory, real-time processing needs, and tight power budgets- MARINERNet’s small size, efficient resource usage, and linear scaling behaviour make it ideal for large-scale deployment. Whether applied to a single radar station or distributed across multiple vessels, the model can scale with minimal additional resource burden, ensuring consistent and efficient performance across different environments. Overall, MARINERNet offers a robust and efficient solution for intrusion detection in maritime radar networks. While the model performs exceptionally well across most classes, further fine-tuning could improve its performance in detecting rarer or more sophisticated attack types, such as Scaling and Translation. Additionally, the model’s performance in real-time environments could be further optimized for low-latency detection, making it highly adaptable for deployment in dynamic and critical radar network environments. Conclusion In this research, we introduced MARINERNet, a novel deep learning-based intrusion detection system designed specifically for maritime radar networks. The study addressed the growing concern of cybersecurity within maritime operations by leveraging deep learning to detect both benign traffic and various cyberattacks such as Denial-of-Service (DoS), scaling, rotation, and addition attacks, among others. Through the use of the RadarPWN dataset, which includes realistic radar network traffic under both normal and adversarial conditions, we demonstrated MARINERNet’s exceptional performance in both multiclass and binary classification tasks, achieving near-perfect accuracy, precision, recall, and F1-scores. The results showcase the model’s robustness, particularly its ability to differentiate between benign and malicious network behavior with minimal misclassification, even under imbalanced data conditions. The significance of these findings lies in MARINERNet’s high potential for real-time deployment in dynamic radar environments, offering a scalable, reliable, and efficient solution to detect intrusions in maritime radar networks. This research underscores the importance of applying deep learning to cybersecurity in critical infrastructure systems, ensuring the safety and resilience of modern maritime operations against increasingly sophisticated cyber threats. Future research in this domain could explore the integration of reinforcement learning (RL) to enable MARINERNet to adapt and evolve with emerging attack strategies in real-time, enhancing its detection capabilities. Additionally, incorporating edge computing technologies could facilitate low-latency, on-site intrusion detection, improving the system’s performance in resource-constrained maritime environments while ensuring faster response times for anomaly mitigation.