Federated Learning Based on Feature Decoupling and Domain Adaptation

Federated Learning (FL) refers to a machine learning approach in which terminal devices collaborate under a server's coordination to collectively refine a global model using their localized data. However, client data heterogeneity can seriously reduce the performance of FL. To tackle the aforementioned issues, this paper presents FedMIDA, a Federated Learning Optimisation Method that utilizes Feature Decoupling and Domain Adaptation. FedMIDA incorporates a feature decoupling module that employs Minimised Mutual Information Neural Estimation (MINE) to eliminate irrelevant information and noise interference from the classification task. It extracts crucial features for adversarial domain adaptation and dynamically adjusts the data distribution between devices to mitigate the negative impact of data heterogeneity. The experimental results demonstrate that the FedMIDA algorithm can enhance accuracy by up to 7.56
^%
compared to other benchmark algorithms under varying degrees of data heterogeneity.