Anomaly Detection in EVCS using Federated Learning with Auto-Optimized Edge Models

This paper presents FedNAS, a novel federated learning (FL) framework designed for robust and personalized models to detect anomalies at distributed electric vehicle charging stations (EVCSs), tailored to real-world industrial deployment. The proposed framework enables collaborative anomaly detection in spatially distributed EVCS without sharing raw data, thus preserving user privacy and complying with data governance requirements. A lightweight neural architecture search (NAS)-optimized Simple Neural Network (SimpleNN) is proposed for each EVCS node to ensure computational efficiency within edge resource constraints. In contrast to traditional methods that depend on fixed structures and simplistic aggregation, FedNAS employs an ensemble-based meta-aggregation strategy, facilitating superior knowledge integration among heterogeneous clients compared to standard Federated Averaging (FedAvg). The framework is evaluated using the IEEE 123-bus test system with multiple EVCS nodes subjected to a cyberattack. The results demonstrate that the FL framework achieves a detection accuracy of up to 98.25%, with significantly reduced communication overhead and improved resilience to adversarial threats. Comparative analysis with centralized learning and federated variants validates the robustness and scalability of the proposed approach outperforms FedAvg by 6.25% in average precision. Furthermore, real-time simulation data is extracted using the OPAL-RT platform to emulate realistic industrial grid conditions, validating the framework's effectiveness in cyberattack detection within dynamic EVCS environments. These findings highlight the potential of federated learning to enable secure, privacy-preserving, and scalable cyberattack detection for next-generation EVCS infrastructure.