enhancing security in federated solar panel fault detection: evaluating robust aggregation methods against malicious client attacks

Renewable energy, particularly solar energy, is critical for sustainable development, yet maintaining solar panel efficiency requires timely and accurate fault detection. federated learning has emerged as a promising solution by enabling decentralized model training while preserving data privacy. however, conventional aggregation methods such as fedavg are vulnerable to adversarial attacks, where malicious clients poison model updates and severely degrade performance. in this paper, we present an enhanced federated transfer learning framework for solar panel fault detection that leverages a pre-trained vgg-16 model for effective feature extraction and incorporates robust aggregation techniques to defend against model poisoning. specifically, we simulate a poisoning attack scenario by introducing malicious clients that inject gaussian noise into their updates, and we evaluate two robust aggregation methods—krum and coordinate-wise median—against this threat. experimental results demonstrate that while standard fedavg yields a final test accuracy of only 19.21% with an exorbitantly high loss, the krum and coordinate-wise median methods achieve significantly improved performance, with final test accuracies of 70.62% and 72.88% and test losses of 1.3161 and 0.7926, respectively. notably, these results are closely aligned with the performance of federated learning without attack, which achieves 74% accuracy with a final loss of 0.82, and centralized learning, which reaches 75% accuracy with a loss of 0.85. these findings underscore the critical importance of robust aggregation in securing federated learning frameworks for solar panel fault detection, providing a scalable and privacy-preserving solution even in adversarial environments.