A Privacy-Preserving Federated Learning Framework for Electric-Vehicle Attack Detection Using Transformers

Abstract

The increasing interoperability between Electric Vehicle (EV) charging networks leads to important cybersecurity challenges, as electric vehicle supply equipment (EVSE) infrastructures play an essential role in the global transformation of clean energy. These systems are increasingly susceptible to security threats for the following reasons: for example, DDoS attack, command spoofing) can cause charging system blackout, and leakage of user privacy information. In this paper, in view of these vulnerabilities, we present a first-party privacy-preserving federated learning (PPFL) system with differential privacy (DP) and Secure Aggregation (SA) , for securing distributed EVSE learning processes. Based on the CIC EV Charger Attack Dataset 2024 (CICEVSE2024), which comprises multimodal data streams (i.e., network traffic, host-based Hardware Performance Counter (HPC) logs, kernel events and power consumption traces), we utilize a transformer-based deep model to capture temporal–spatial correlations across different modalities for anomaly detection. We systematically compared the four FL configurations (standard FL, FL + DP, and corresponding to FL + Secure Aggregation and finally FL+DP+Secure Aggregation) from the perspective of utility-privacy-efficiency trade-off. The results show that the hybrid FL + DP + SA model reduced privacy risk by 20%, and decreased membership inference attack (MIA) success rate from 55 to 35%, while still maintaining more than 70% detection accuracy with only a slight increase (30–40%) in convergence time comparing with standard FL. These results validate that the proposed design is a well-architected, scalable, and robust solution for protecting the next-generation EVSE infrastructure against distributed cyber attacks and privacy attacks.