Non-IID Free Federated Learning With Fuzzy Optimization for Consumer Electronics Systems

Consumer electronics suffer from data surges which lead to data privacy and integrity issues. Federated Learning (FL) keeps data with local devices which enhances privacy, and efficient device collaboration by processing data locally without transmitting sensitive information to centralized servers. However, data heterogeneity within practical FL settings poses a formidable challenge, such as client model drift, significantly impacting the convergence speed of the global model and leading to a substantial communication overhead. To address these issues arising from data heterogeneity, we propose a novel FL framework based on latent dependency optimization with fuzzy-based client selection (FCS), named FedLaDO. Specifically, in the consumer electronic client-side training, we introduce two new loss regularizers: a positive latent dependency loss and a perceptive knowledge co-distillation loss. These regularizers are integrated into the client training objective function to enhance the client models’ reinforcement learning over its distinctive training data while preserving the global knowledge inherited from the prior global model. For consumer electronic server-side model aggregation, we propose an FCS mechanism to filter out underperforming and overperforming clients and introduce a novel inverse dependency aggregation algorithm to effectively integrate drifted client models. The efficacy of our approach is validated through extensive experiments on diverse real-world datasets, showcasing its superior performance in achieving higher model convergence accuracy compared to state-of-the-art FL baseline methods.