Robust Federated Learning With Multi-Category Accuracy for Non-IID Data

Federated learning is increasingly utilized in devices such as smartphones, smart home systems, and wearables for applications like personalized recommendations, voice assistants, and health monitoring. However, FL’s decentralized nature makes it susceptible to malicious attacks, and the inherent non-independent and identically distributed (non-IID) data in consumer electronics further complicates the identification of malicious actors. Existing research primarily defends against these attacks by clustering and aggregating gradient information from all clients, reducing the influence of malicious actors. Nevertheless, these methods often overlook their effectiveness in the context of non-independent and identically distributed (non-IID) data, which makes it difficult to identify malicious attackers under such data distributions. In non-IID scenarios, the gradients from clients are often diverse, and this variability can obscure the presence of malicious updates, as they may blend with naturally occurring anomalies in the data. To address these issues, we propose a robust Federated Learning framework for security attacks and heterogeneous data, based on multi-category accuracy, to distinguish the gradients submitted by regular participants from those of malicious attackers. This framework is named RMCA-FL. Finally, we conduct extensive experiments to evaluate the performance of the proposed framework by comparing it with other standard algorithms. The experimental results demonstrate that the global model with RMCA can guarantee an accuracy of about 97.1%, compared to 61.0%, 59.4%, and 50.9% for Fedavg, MWU, and MWU*, respectively.