Personalized and similarity contrast federated learning on non-IId

Data heterogeneity is a major challenge in federated learning. This paper introduces FedKSA, a personalized approach that extends FedKD to handle non-iid data. FedKSA integrates an adaptive local aggregation module in client models for dynamic global and local updates tailored to each client's objectives. It uses gradient-based clustering during model aggregation to reduce data distribution differences, enhancing performance. Experiments across diverse datasets and environments show FedKSA effectively mitigates data heterogeneity, outperforming original and baseline methods with up to 1.84% and 45.22% increases in test accuracy, respectively.