Targeted Adversarial Poisoning Attack Against Robust Aggregation in Federated Learning for Smart Grids

Federated learning has been widely adopted as an intelligent computing method to leverage the power of distributed data without compromising data privacy and security, especially in smart grids. Poisoning attack is one of the common attacks against federated learning to degrade the performance of the global model or cause it to make specific, incorrect predictions. To counter these threats, secure aggregation rules have been implemented to reduce the impact of adversarial or malicious updates during training process. In this paper, we first propose a norm-based aggregation rule specifically designed to mitigate the effects of poisoning attacks within federated learning systems used for power quality classification. Subsequently, we introduce the Targeted Adversarial Poisoning Attack (TAPA), a method that aims to break the secure aggregation rule by causing the global model to misclassify a specific targeted class as another chosen class with increased success. This attack method employs a generative approach to produce extra fake data and trains a malicious model whose parameter norms closely resemble those of standard local models, thereby increasing the likelihood of these updates surviving the filtering processes of secure aggregation rules. Evaluation results demonstrate that the proposed TAPA method can achieve a higher poisoning attack success rate while preserving the classification accuracies of non-targeted classes, even in the presence of unbalanced data distributions. This approach underscores the potential vulnerabilities in existing secure aggregation mechanisms and highlights the need for robust defenses against sophisticated poisoning strategies in federated learning environments.