Toward Privacy Preservation in Federated Learning: A Framework Integrating Client-Side Shuffling and Model Compression

Federated learning (FL) enables multiple clients to train models on local data and collaboratively optimize a global model without sharing raw data. However, client heterogeneity, such as differences in data distributions and system capabilities, poses challenges like reduced training efficiency and slower convergence. Furthermore, there is a risk of inference attacks during the transmission of model parameters. To address these issues, we propose the client-side shuffling and compressed-model FL (CSCP-Fed) framework, which is efficient and privacy-preserving, based on client-side shuffling and compressed-model techniques. The framework combines differential privacy (DP) with client-side shuffling to ensure data anonymity and enable secure weighted aggregation of model parameters. It also designs an asymmetric-encryption-based secure communication protocol to safeguard data transmission. In addition, it introduces a hybrid-weighted attention aggregation algorithm and a compressed-model-driven client selection strategy to mitigate the impact of heterogeneity, accelerate convergence, and maintain model generalization ability. Rigorous security analysis and experiments show that CSCP-Fed can effectively protect privacy without relying on any centralized entity. It reduces the expected global loss per round by 15%-38%, cuts communication overhead by 19.3%-44.3%, speeds up convergence by 10%-41%, and improves learning accuracy by 4%-13% compared to traditional methods.