Interpretability-Based Cross-Silo Federated Learning

The severe challenge encountered in cross-silo federated learning (FL) is the performance degradation caused by data heterogeneity. To overcome it, we propose two methods, FedGDI and FedCI, identifying these clients with unbalanced categories based on an interpretability mechanism. We firstly iteratively generate feature maps of last global model and local client models selected, then these unbalanced local models are identified by comparing the feature maps. For clients with unbalanced categories, these local update parameters are further adjusted by minimizing the gradient distance between the global model and clients’ models, so as to reduce the adverse impact on the performance of FL model. We adopt different client-filtering strategies to filter clients. FedGDI filters clients by taking advantage of the cosine similarity between the gradient of these local client models and last aggregation global model; FedCI samples clients by clustering clients based on gradient distance. We evaluate the effectiveness of FedGDI and FedCI through multiple datasets, and from experimental results it can be concluded that our methods outperform these state-of-the-art (SOTA) schemes.