Federated Learning, Temporal Convolutional Networks, Electric Vehicle Charging Infrastructure, Intrusion Detection, Cybersecurity, Privacy-Preserving Machine Learning

Abstract The rapid growth of electric vehicle (EV) charging infrastructure has increased exposure to cyberattacks, while conventional centralized intrusion detection remains poorly suited to distributed and privacy-sensitive EVSE environments. This study proposes a federated learning framework based on a Dual-Attention Temporal Convolutional Network (DA-TCN) for collaborative cyberattack detection without centralized data sharing. The model combines dilated causal convolutions with channel and temporal attention to capture complex dependencies in multimodal EVSE telemetry. To improve federated optimization, we introduce Federated Stochastic Weight Averaging (FedSWA) and client-local mixup augmentation, which together enhance generalization and robustness across distributed clients. Evaluated on the CICEVSE2024 dataset, the proposed framework achieved 99.18\% accuracy and 98.95\% macro F1-score, outperforming a centralized baseline (98.96\% accuracy) while maintaining perfect recall for benign and cryptojacking traffic. With an inference latency of 7.21~ms per batch and a model size of 4.07~MB, the framework is well suited for real-time edge deployment. A complementary Isolation Forest module provides unsupervised zero-day anomaly detection. These results demonstrate that federated learning can deliver accurate, lightweight, and deployment-ready intrusion detection for privacy-preserving EV charging infrastructure.