Securing Smart Grid Federated Learning Against Advanced Evasion Attacks Using Ensemble-Based Adversarial Training

Smart grid Advanced Metering Infrastructure (AMI) integrates smart meters, communication networks, and data systems to enhance electricity monitoring and control. Smart meters provide real-time consumption data for accurate billing and energy management, but AMI is vulnerable to false data injection attacks, where consumers manipulate meters to lower their bills, leading to financial losses and grid instability. Machine learning-based detection methods are effective but they require large, diverse datasets, yet electrical utilities often lack sufficient data and are reluctant to share due to privacy and legal concerns. Federated learning enables decentralized model training without data sharing but it is vulnerable to evasion attacks—where adversaries subtly modify malicious data to bypass detection. While adversarial training is effective in centralized settings, applying it in federated learning is challenging. Unlike centralized systems, where a trusted entity enforces security, federated learning distributes data across multiple clients, making it unrealistic to assume all participants follow security protocols. Malicious clients can exploit this decentralized nature to bypass defenses and inject vulnerabilities into the global model. This paper introduces advanced attacks in which malicious clients generate adversarial samples using different techniques than honest clients, significantly increasing attack success rates. Using real-world electricity consumption data, we evaluate the impact of these attacks and propose a novel defense mechanism that incorporates ensemble learning, adversarial training, and median-based filtering. Extensive experiments demonstrate that our approach effectively strengthens federated learning-trained models against evasion attacks, enhancing smart grid AMI security.