CSPFed: Compressed Sensing-Enhanced Privacy-Preserving Federated Learning for Efficient Multi-Domain Scenarios

Federated Learning enables collaborative model training across multiple devices without sharing raw data. However, when private data originates from multiple domains with different feature distributions, the model's generalization ability on unseen domains can deteriorate. Moreover, existing methods often overlook privacy protection and communication overhead during model performance enhancement. To address these challenges, we propose the CSPFed Framework. First, we introduce Federated Prototype Learning to improve the generalization of the Global Model on private domains. Next, we employ improved Compressed Sensing for lossy compression of the model. Finally, we propose a block-based CKKS Homomorphic Encryption Algorithm to encrypt data transmitted between clients and the server. This framework significantly reduces communication overhead while enhancing model accuracy, all while ensuring data privacy. CSPFed outperforms the best competitors by 5.59 % on two public datasets. Empirical results demonstrate the effectiveness of the proposed solution and the efficiency of its key components.