Adversarially Robust Deep Learning Models for Secure Cyber-Physical Systems: A Multi-Layer Defense Framework for Real-Time Threat Mitigation

Cyber-physical systems (CPS) are increasingly dependent on deep learning models for real-time monitoring, control, and decision-making, making them vulnerable to adversarial attacks that can compromise both computational and physical operations. This study investigates the development of adversarially robust deep learning models integrated within a multi-layer defense framework to enhance security and ensure real-time threat mitigation in CPS. A quantitative experimental approach was employed, using Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks trained on clean and adversarially perturbed datasets generated using Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). Data preprocessing, feature extraction, and adversarial training were combined with system-level defenses such as secure communication, redundancy mechanisms, and anomaly detection to form a comprehensive multi-layer framework. Statistical analysis, including ANOVA and regression, was conducted to evaluate performance and robustness across various experimental setups. Results indicate that standard CNN and LSTM models achieved high accuracy on clean data (CNN: 96.2%, LSTM: 94.7%) but were highly vulnerable to adversarial attacks (CNN: 68.5%, LSTM: 86.1%). Incorporating adversarial training and multi-layer defenses improved robustness significantly, with accuracy above 90%, F1-scores above 0.90, detection rates up to 96.1%, and latency maintained below 30 ms, demonstrating reliable real-time protection. These findings confirm that a holistic, layered security approach effectively mitigates adversarial threats, preserves operational integrity, and supports the deployment of trustworthy AI-driven CPS.