Abstract: Combining vehicle-to-everything (V2X) communication with smart grid technology enables intelligent electric vehicle management, dynamic charging, demand response, and cyber-physical monitoring in real time. Nevertheless, such a complex environment faces privacy leakage, data injection falsification, spoofing, denial of service, poisoning, evasion, and gradient inference attacks. Although existing federated learning-based approaches reduce the amount of raw data that must be shared, they apply fixed differential privacy noise and make assumptions about the uniformity of client reliability. Furthermore, existing frameworks show little resilience to poisoning, evasion, spoofing, data injection, and denial-of-service/DDoS attacks owing to their inability to work properly with non-IID vehicles and electric vehicle data. To solve the abovementioned problems, this study proposes a Trust-Aware Adaptive Differentially Private Hierarchical Federated Adversarial Learning framework, which employs a Convolutional Neural Network-Graph Convolutional Network-Gated Recurrent Unit (CNN-GCN-GRU), for intrusion detection in V2X-enabled smart grids. The architecture uses a three-layer approach with vehicle and electric vehicle clients, edge aggregators that can be either roadside units or charge stations, and a smart grid cloud center. The CNN component extracts local traffic data, the GCN models the topological dependencies between the vehicle and grid nodes, and the GRU component captures attack-related behavior patterns. Adaptive differential privacy takes advantage of the trust scores of different clients to adjust the amount of noise applied, whereas trust-aware aggregation is utilized to lower the contribution from potentially unreliable updates. Finally, adversarial learning is employed to improve resilience to the aforementioned attacks. Non-Independent and Identically Distributed (Non-IID) settings are evaluated using metrics such as accuracy, precision, recall, F1-score, false alarm rate, privacy budget, communication overhead, detection delay, reliability, scalability, and resilience. Based on the experimental results, the TADP-HFAL algorithm obtains an accuracy of 97.20%, an F1-score of 96.67%, a false alarm rate of 2.85%, and reduces communication overhead by 51.59% compared to the traditional flat federated learning approach.

Key words: intrusion detection, cyber-physical networks, privacy leakage, non-IID data distribution, federated learning, false data injection