FedCA: Federated learning based on classification layer alignment

Gang Liu, Jing Jia, Rui Mao, Yan Ji

Abstract: The issue of data heterogeneity has long been a significant challenge in federated learning, impacting the speed of global model convergence and potentially lowering the overall performance of the global model, thereby impeding its development. In response to this challenge, we propose a classifier head-based method to correct the local model update direction. Unlike most existing calibration methods, our method not only ensures a global model with good performance but also guarantees the performance of a single local model. Concretely, we have introduced a new federated learning framework called FedCA. In this framework, the output of the feature extractor is imported into both a global classifier and a local classifier. The gradients are calculated separately, and then backpropagate into the model. This method ensures the performance of the local model and allows the client to maintain a consistent learning direction. Our experimental results demonstrate the feasibility of this method.

External IDs:dblp:conf/icmcs/LiuJ0J24