Adaptive Intrusion Detection in Cyber-Physical Systems Using Reinforcement Learning-Based Autoencoders

The increasing integration of Cyber-Physical Systems (CPS) in critical infrastructure presents unique challenges for ensuring robust cybersecurity. Traditional Intrusion Detection Systems (IDS) often struggle with the dynamic, complex nature of CPS environments, making adaptive solutions essential. This paper presents an Adaptive Intrusion Detection System (AIDS) that leverages Reinforcement Learning-Based Autoencoders (RL-AE) to detect and respond to cyber intrusions in CPS environments. The proposed model combines the power of autoencoders for unsupervised anomaly detection with reinforcement learning to continuously adapt and improve detection accuracy. Experimental results on benchmark CPS datasets demonstrate that the RL-AE-based IDS achieves a detection accuracy of 98.3%, a precision rate of 96.8%, and a recall rate of 97.5%, effectively minimizing false positives to 2.4%. The system also demonstrated rapid adaptability, with a convergence time of less than 50 iterations in response to new attack patterns. These results highlight the model's ability to enhance system resilience and promote secure and efficient CPS operations in industrial and critical infrastructure applications. This approach provides a scalable and flexible IDS solution, promoting secure and resilient CPS operations in real-time, high-stakes environments.