Evolutionary cross-client network aggregation for personalized federated learning

Personalized federated learning (PFL) enables the training of machine-learning models on dispersed data while ensuring user privacy. Despite their advantages, existing PFL algorithms primarily focus on model architectures and communication issues and often neglect the crucial aspects of efficiency and adaptation in addressing problems trapped in local optima. In this study, we propose an innovative solution to overcome this limitation by introducing an evolutionary cross-client aggregation approach. Our approach aims to enhance personalized model performance in non-independent and identically distributed (non-IID) FL settings. By leveraging the diversity of the client models and employing mutation-based cross-client aggregation, we generate new architectures that are subsequently trained locally during the FL process. We evaluated the effectiveness of our proposed method through experiments on a benchmark dataset and compared its performance with that of other FL methods. The results demonstrate the superiority of our approach, which achieved an impressive accuracy of 85% and outperformed existing methods.