Deep Learning Models for Predicting Cyber‑Physical Attacks in Supply Chain Networks

Abstract

Cyber-physical supply chain networks have become increasingly vulnerable to sophisticated, multi-stage cyber threats due to the convergence of operational technology, enterprise systems, and real-time IoT infrastructures. Traditional rule-based and reactive security mechanisms are often insufficient to detect stealthy or evolving attack strategies that propagate across interconnected nodes. This study proposes a spatial-temporal deep learning framework for predictive cyber-physical attack detection in supply chain environments. The model integrates Graph Convolutional Networks to capture structural interdependencies between distributed supply chain entities and Long Short-Term Memory networks to model sequential behavioral anomalies over time. The proposed architecture generates probabilistic early-warning signals distinguishing normal operations, pre-attack anomalies, and active attack states. Experimental evaluation against conventional baselines demonstrates improved classification performance and reduced detection latency. The results highlight the effectiveness of hybrid graph-based and temporal learning approaches in shifting supply chain cybersecurity from reactive detection toward anticipatory risk modeling.