With Cyber-Physical Systems (CPS) ranging from smart grids and self-driving cars to healthcare networks becoming the cornerstones of infrastructure in today’s world, detecting anomalies in real-time is crucial to ensure safety, guarantee reliability, and preserve performance. Traditional supervised learning methods face challenges in real-world CPS due to the limited availability of labeled data. Even when labels exist, they often become outdated because of sensor drift, system reconfigurations, or new types of attacks. This survey presents a comprehensive review of recent self-supervised learning (SSL) methods for anomaly detection in CPS, categorized into contrastive, reconstructive, predictive, and joint embedding-based paradigms. We further summarize hybrid SSL models that integrate SSL with generative modeling, domain adaptation, federated learning, and meta-learning to enhance adaptability and deployment feasibility. Drawing from reported results on real-world CPS datasets, we benchmark 29 methods using standard metrics such as F1-score and AUROC, while noting the limitations of commonly used CPS datasets. To the best of our knowledge, this is the first survey that (1) systematically reviews both traditional and hybrid SSL approaches tailored to CPS, (2) identifies Joint Embedding Predictive Architectures (JEPA) as an emerging fourth category alongside contrastive, reconstructive, and predictive methods—an inclusion often overlooked in traditional SSL categorizations, and (3) benchmarks recent SSL methods across a wide range of CPS datasets. These contributions offer a unified foundation for researchers and practitioners aiming to develop robust, adaptive, and label-efficient anomaly detection systems in evolving CPS environments.

Drones pose significant security threats due to their stealth and versatility. Brushless DC (BLDC) motors used in drones emit unique electromagnetic signals useful for drone detection, but environmental noise often degrades their quality and interpretability. This paper presents a robust Generative Adversarial Network (GAN) based framework to enhance signal clarity through denoising. Trained on paired noisy and clean signals obtained from two drone motor types (namely A and B), the GAN outperforms traditional methods (BM3D, Wavelet, Wiener, and EMD), achieving a 30.65% and 33.04% reduction in Mean Squared Error (MSE) for motors A and B, respectively. The average signal-to-noise ratio (SNR) is increased by 1.58 dB for motor A and 1.75 dB for motor B. The GAN model is then evaluated using a convolutional neural network (CNN) classifier trained on spectral correlation density (SCD) images, and it demonstrated substantial improvements in accuracy, precision, recall, and F1-scores compared to baseline results obtained from untreated signals. Specifically, classification accuracy improved significantly, from 76.38% to 98.67% for Motor Type A (29% improvement) and from 81.97% to 97.27% for Motor Type B (19% improvement), resulting in an average improvement of 24%. These results highlight the GAN-based denoising strategy’s ability to enhance signal interpretability and diagnostic reliability, significantly advancing drone motor detection capabilities.

Ground segmentation from LiDAR point cloud data plays a critical role in both civil engineering and autonomous vehicle systems. However, real-world LiDAR data often suffers from geometric distortions, occlusions, and dense clutter, which limit the reliability and accuracy of ground segmentation. To overcome these challenges, we introduce KPU-Net. This deep neural network architecture employs: 1) T-Net module, which handles geometric distortions by aligning point clouds into a canonical pose, 2) KPConv-augmented U-Net encoder-decoder, which handles occlusions, clutter, and irregular terrain by capturing fine-grained, hierarchical features through learned kernel point convolutions over local neighborhoods. In addition to above, KPU-Net offers following advantages: 3) KPU-Net features high speed processing (approximately 231K points per second making it well-suited for scalable deployment in mapping and perception systems), while 4) preserving the original point cloud density (i.e no loss in point cloud data, avoiding sparsification that can compromise precision in various applications). The framework was trained and tested on a benchmark dataset and diverse point cloud data collected by the Timmons group, covering urban, vegetation, and complex terrain environments. The presented KPU-Net was evaluated against five widely used LIDAR data segmentation methods: Random Forest, PointNet, GndNet, RandLA-Net, and KPConv. The proposed KPU-Net demonstrated better performance on mean Intersection over Union (mIoU, up to 33%), mean accuracy (mAcc, up to 25%), and overall accuracy (OA, up to 19%), over the five compared state-of-the-art methods.

The Internet of Vehicles (IoV) has transformed intelligent transportation systems through vehicle-to-everything (V2X) communication, improving road safety and traffic efficiency. However, the dynamic nature of vehicular networks, with high mobility and shared wireless resources, makes them vulnerable to attacks like Denial of Service (DoS). Anomaly detection (AD) has proven effective in detecting such threats. Yet, V2X communication occurs in diverse environments with varying network coverage and vehicle speeds, leading to domain shifts and variations in feature distributions that can hinder the generalization performance of traditional anomaly detection models. To address these challenges, this paper presents V2XFormer, an unsupervised anomaly detection system based on transformer neural networks, designed to identify anomalies in V2X communication. Additionally, we introduce TV2XFormer, which integrates transfer learning to enhance adaptability across diverse network conditions and environmental variations in V2X communication. We assess the performance of the proposed approaches using the VDoS-LRS V2X dataset, employing precision, recall, and$\mathbf{F 1}$score metrics. A comparison is made with five state-of-the-art unsupervised AD algorithms. Experimental results demonstrate that both V2XFormer and TV2XFormer outperform the competing algorithms, achieving the highest$\mathbf{F 1}$scores (1.0). Furthermore, TV2XFormer exhibits notable robustness and generalizability to dynamic vehicular environments.

As Deep Learning (DL) algorithms become more widely adopted in healthcare applications, there is a greater emphasis on understanding and addressing the potential privacy risks associated with these models. The purpose of this study is to investigate the privacy vulnerabilities of the Convolutional Neural Network (CNN) classifiers for Electroencephalogram (EEG) data in the Brain-Computer Interfaces (BCIs). Specifically, it focuses on the Membership Inference Attack (MIA), which seeks to determine if data from an individual were used in model training. The novelty of this work lies in its empirical analysis of MIA, by addressing two key challenges that are less common in other domains: 1) heterogeneous datasets and 2) spatio-temporal design choices. Motivated by these challenges, we investigate the susceptibility to MIA based on: 1) the specifics of the training data set (number of participants, demographics), and 2) specifics of the CNN (such as architecture, regularization). Our experiments revealed that an adversary with limited knowledge of the model and its training process can compromise the privacy of training participants, noting that the same attack is not effective against deep learning models trained on image and tabular datasets. Some of our findings are: 1) training on diverse participant datasets improves the privacy of most participants but increases risks of memorization and vulnerabilities for underrepresented groups; 2) regularization is less effective in defending against the MIA on EEG data CNN classifiers when compared to other types of input data; 3) the depth and width of the model architecture have no impact on the effectiveness of membership attack. We hope that the insights presented will help future researchers develop more privacy-aware deep learning-based BCI systems.

The advancements in Cyber-Physical Systems (CPSs) have also increased their vulnerability to various cyber-attacks. Therefore, it is crucial to develop strong cybersecurity mechanisms, shielding these critical systems from potential cyber intrusions. Among many AI technologies, Generative AI (GenAI) has gained significant attention in the last couple of years. This is due to its distinctive capability to autonomously generate original and diverse content across different domains, offering potential for novel advancements in several applications. Given the massive success of GenAI, it is essential to explore its role in ensuring the cybersecurity of CPSs. Therefore, in this paper, we present: 1) the evolution and current state of GenAI, 2) benefits of GenAI on the cybersecurity of CPS, 3) threats of GenAI on the cybersecurity of CPS, 4) defense strategies against threats and 5) future research opportunities. We hope this systematic survey will help the community prioritize research efforts to address pressing issues in cybersecurity of CPSs.

With the many advancements in automobile technology, there has been a sharp increase in the number of sensors and systems present in vehicles. This has enabled a rapid increase in the features and capabilities available in modern automobiles but has also vastly increased their vulnerability surface. Attackers are now able to remotely attack and control some facets of modern automobiles, which creates dangerous situations for drivers and passengers. In order to address this, this paper proposes a self-attention bottleneck network utilizing an encoder-decoder architecture. This is used as an anomaly detection system (ADS) that can detect anomalous behavior present within CAN bus communications. We evaluated this approach using a publicly available CAN bus car hacking dataset and show that our architecture is able to achieve an accuracy of over 99% for detecting anomalies present in CAN bus data.

Electric Vehicle Charging Stations (EVCS) are rapidly being built all around the world to support the growing number of Electric vehicles (EVs) on the road. EVCSs hook into critical infrastructure and offer a vital service, so it is very important they remain secure and available, yet they remain vulnerable to a number of attacks. Thus, it is becoming more and more important to examine the security of EVCSs. In this paper, we will explore the current cyber-physical threat landscape faced by EVCSs. First, we examine the protocols used for communication during the EV charging process and their strengths and weaknesses. Next, we discuss the overall cyber-physical security threats of the entire system, as well as Artificial Intelligence-based (AI) solutions to combat these threats. Finally, we present the future research directions.

With the advancements of modern vehicle infrastructures, vehicles are increasingly relying on the signals received from a vast number of sensors and electronic components. Wireless technologies enable communication between vehicles and infrastructure, but it also increase the vulnerability surface. Malicious actors can remotely disrupt the vehicle's normal behavior, causing vehicle damage or worse, putting human lives in danger. To address these challenges, this paper proposes a transformer neural network-based intrusion detection system (CAN-Former IDS) that predicts anomalous behavior within the CAN protocol communication. Previous work typically addresses the prediction over the sequence of the CAN IDs. In this paper, we will simultaneously analyze both the sequence of IDs and the message payload values. The advantages of our approach are: 1) fully self-supervised training, which does not require labeled data, 2) self learning interactions between input tokens without relying on hand-crafted features. The transformer neural network is trained to predict the next communication sequence and anomalous communication is identified by comparing the real sequence to the predicted expected sequence. We evaluated our approach using a publicly available data set known as survival analysis data set, containing CAN communication from three different cars.

Modern vehicles have multiple electronic control units (ECUs) that are connected as part of a complex cyber-physical system (CPS). The controller area network (CAN) is a well-known communication protocol that connects these ECUs because of its reliability and efficiency. However, adversaries can easily inject abnormal messages into the CAN bus remotely to affect vehicle driving safety. Existing anomaly detection methods only focus on specific vehicle models and have a limited range of applications across different vehicles. To address this challenge, this paper proposes a Domain Adversarial training-based AutoEncoder (DAdAE) for unsupervised CAN anomaly detection. The advantages of our approach are: 1) detect variant attack scenarios on different car models 2) does not require labeled data 3) works well even with a limited dataset. The effectiveness of the proposed model is evaluated on the survival dataset, and the experiment results show that the DAdAE model improves the overall f1 score significantly, compared to other unsupervised models.

Recent year has brought considerable advancements in Electric Vehicles (EVs) and associated infrastructures/communications. Intrusion Detection Systems (IDS) are widely deployed for anomaly detection in such critical infrastructures. This paper presents an Interpretable Anomaly Detection System (RX-ADS) for intrusion detection in CAN protocol communication in EVs. Contributions include: 1) Feature Extractor; 2) Anomaly Detection System; and 3) Explanation Generator for detected anomalies. The presented approach was tested on two benchmark CAN datasets: OTIDS and Car Hacking. The anomaly detection performance of RX-ADS was compared against the state-of-the-art approaches on these datasets: HIDS and GIDS. The RX-ADS approach showed comparable performance to the HIDS approach on OTIDS dataset and outperformed HIDS and GIDS approaches on Car Hacking dataset. Further, the proposed approach was able to generate explanations for detected abnormal behaviors arising from various intrusions. These explanations were later validated by information used by domain experts to detect anomalies. Other advantages of RX-ADS include: 1) the method can be trained on unlabeled data; 2) explanations help experts in understanding anomalies and root course analysis, and also help with AI model debugging and diagnostics, ultimately improving user trust in AI systems.

In critical infrastructures, timely detection of anomalies is essential to detect failures, avoid catastrophic damages, and improve resilience. Neural Network models are one of the state-of-the-art approaches used for anomaly detection. Among Neural Network architectures used these days, Autoencoders (AEs) have gained significant attention due to their advantages such as unsupervised learning, dimensionality reduction, non-linear feature extraction, the ease of integration with other neural network algorithms, and ease of use. Therefore, in this paper, we present: 1) anomaly detection and types of anomaly detection, 2) recent advancements in AEs typically used in anomaly detection, 3) AE-based Anomaly Detection (AE-AD) in selected critical infrastructures such as smart grids, intelligent transportation systems, and smart buildings, and 4) future research opportunities. We hope that this systematic survey of AE-based anomaly detection approaches will help the community prioritize research efforts to address pressing issues in critical infrastructures.