PPCSA: Partial Participation-Based Compressed and Secure Aggregation in Federated Learning

Federated Learning (FL) enables users devices (UDs) to collaboratively train a Deep Learning (DL) model on an individual’s gathered data, without revealing their privacy sensitive information to the centralized cloud server. Those UDs usually have limited data plans with a slow network connection to a centralized cloud server, which causes limited communication bandwidth between the contributing mobile users. To mitigate this problem, we propose a novel Partial Participation-based Compressed and Secure Aggregation (PPCSA) algorithm. To implement the PPCSA, we use a Sparse Compression Operator (SCO) that reduces the communication bits between the cloud server and the users while maintaining the FL requirements. In particular, PPCSA utilizes a novel compression method and introduces a Local Differential Privacy (LDP) based framework to achieve the communication-efficiency at a new level. Our experiments on a commonly used FL dataset show that PPCSA distinctively outperforms the state-of-the-art schemes in terms of convergence accuracy and communication bits.