Multimodal Anomaly Detection for Autonomous Cyber-Physical Systems Empowering Real-World Evaluation

As autonomous Cyber-Physical Systems (CPS) increasingly operate in critical environments, ensuring their security and reliability becomes paramount. This paper presents a robust anomaly detection framework designed to enhance the resilience of CPS by integrating multiple sensor modalities, including Lidar, Odometry, and Network Traffic. Our approach leverages the strengths of each modality, compensating for potential weaknesses when individual modalities are considered in isolation. A vector-based reconstruction loss function is introduced, significantly improving the detection of subtle anomalies by preserving the contributions of individual features. Our experimental evaluation, conducted on a custom-built Unmanned Ground Vehicle (UGV) platform, shows that the proposed system achieves an anomaly detection accuracy of up to 98% when using the improved vector-based reconstruction loss, compared to 72% with a standard scalar-based loss. Even when the training data is reduced by 50%, bringing the total training set size down to 92 samples, the system maintains a high accuracy of 97%, demonstrating its robustness under constrained data conditions. These results indicate the effectiveness of our multimodal approach in real-world applications where data availability may be limited. Our work focuses on generalizability and modularity, ensuring adaptability across various CPS platforms and evolving threats, ultimately enhancing the reliability of autonomous systems in real-world scenarios.