Adaptive model poisoning detection in federated learning for EV charging networks: a blockchain-anchored explainable AI approach

Abstract

The rapid expansion of electric vehicle charging infrastructure necessitates collaborative machine learning approaches to detect cyber threats across distributed networks while preserving operator privacy. Federated learning enables decentralized threat intelligence sharing among charging station operators without exposing sensitive operational data. However, this distributed paradigm introduces critical vulnerabilities to model poisoning attacks, where malicious participants inject corrupted updates to degrade global model performance or introduce backdoors. This paper presents a novel adaptive detection framework that combines blockchain-based integrity verification with explainable artificial intelligence to identify and mitigate poisoning attempts in federated learning systems for electric vehicle charging networks. The proposed architecture employs a multi-layered defense mechanism incorporating statistical anomaly detection, gradient-based behavioral analysis, and blockchain-anchored audit trails to ensure model update authenticity. An explainability module utilizing Shapley values provides interpretable insights into detected anomalies, enabling human-in-the-loop validation and regulatory compliance. Experimental evaluation using a synthesized electric vehicle charging network dataset with simulated poisoning attacks demonstrates detection accuracy exceeding 94 percent across various attack scenarios including label flipping, gradient manipulation, and backdoor injection under controlled simulation conditions. The framework maintains federated learning convergence while introducing minimal computational overhead of 9.3 percent compared to baseline implementations. Results indicate the approach successfully balances security, privacy, and operational efficiency requirements for critical energy infrastructure applications.