Feature Partitioning for Federated Learning with Momentum-Guided Aggregation

Federated learning allows edge devices to collaboratively train deep models without uploading raw data, but one of its main challenges is the presence of non-independent and identically distributed (non-IID) data. Specifically, the varying data distributions across different clients result in slower convergence and reduced model accuracy during training. To confront this challenge, this paper proposes a method that partitions the features of raw data into sensitive and robust features. The sensitive features are used to generate noisy shared data, which is then uploaded to the server. The server trains on these pseudo images, while clients perform local training using the original data. Model parameters are aggregated and exchanged between the server and clients at predefined communication intervals. In addition, we adopt a momentum-guided initialization strategy for parameter aggregation. The experimental results show that the proposed method outperforms the state-of-the-art methods by $\mathbf{5. 3 1 \%}$ in terms of performance.