CosPer: An adaptive personalized approach for enhancing fairness and robustness of federated learning

Federated learning (FL) enables clients to collaboratively train a global model while safeguarding the privacy of their respective data. In practical applications, the data heterogeneity across clients introduces competing constraints on the accuracy, fairness and robustness of FL models. However, existing methods that attempt to address all three of these issues perform poorly in scenarios with severe data heterogeneity. To address this limitation, this paper proposes an innovative personalized federated learning (PFL) framework named CosPer. This framework employs an adaptive local aggregation mechanism, where the aggregation weight is determined by the cosine similarity between the global gradient and the local gradient of each client. This mechanism can construct a personalized model for each client, suited to different degrees of the data heterogeneity. In addition, we conduct extensive experiments to evaluate the performance of CosPer on four datasets. The results demonstrate that CosPer outperforms other state-of-the-art PFL methods in terms of accuracy, fairness and robustness.