FEAR-DPTGAE: Federated Explainable Adversarially Robust Differentially Private Temporal Graph Autoencoder for Secure Smart Transportation

Abstract

Smart transportation systems produce vast amounts of dynamic graph-structured data from vehicles, roadside units, IoT sensors and V2X communication networks, posing crucial challenges related to privacy protection, adversarial resilience, explainability, and scalable federated intelligence. Current temporal graph learning systems solve these problems one at a time, leading to fragile solutions that fail in face of a real-world mix of threats. It is proposed that a unified Federated Explainable Adversarially Robust Differentially Private Temporal Graph Autoencoder (FEAR-DPTGAE) is a Federated Explainable Adversarially Robust Differentially Private Graph Autoencoder with four novel mechanisms: Federated Explainable Adversarially Robust Differentially Private Temporal Graph Autoencoder - Adaptive Temporal Graph Privacy Allocator (ATGPA): An entropy-guided framework on budget allocation per node, Federated Explainable Adversarially Robust Differentially Private Temporal Graph Autoencoder - Cross-Client Adversarial Robustness Aggregation (CARA): A protocol for Byzantine-resilient federated training, Federated Explainable Adversarially Robust Differentially Private Temporal Graph Autoencoder - Graph Attention Explainability Regularizer (GAER): A regularizer to provide consistent node attribution under privacy noise for semantically consistent graph, Federated Explainable Adversarially Robust Differentially Private Temporal Graph Autoencoder - minimax-optimal adversarial temporal graph training objective. Formal (ε,δ)-DP composition proofs are given. The evaluations on METR-LA, PEMS-BAY, VeReMi, and TDrive datasets show that FEAR-DPTGAE achieves a high anomaly detection accuracy of 96.1% at ε=1.0 with a low traffic prediction MAE of 2.14 km/h, outperforming all baselines and maintaining high Explainability Score (0.92) while being robust against up to 30% malicious clients.