Privacy-Preserving and Efficient Model Aggregation in Edge-Assisted Federated Learning

Integrating edge computing technology with federated learning (FL) can significantly enhance model training and communication efficiency compared to the traditional cloud-client architecture. However, semi-asynchronous edge-assisted FL still faces security challenges, particularly privacy breaches resulting from transmitted parameters. To address this issue, we propose an efficient and secure framework for semi-asynchronous edge-assisted FL. We leverage a supervised differential privacy (DP) technique employing control variables to ensure efficiency and security during the aggregation process. The control variables effectively guide the model training process, thereby mitigating loss of accuracy. We also consider practical applications and propose an algorithm to minimize the impact of client drop-out. The results of the experiments confirm the practicality and effectiveness of our protocol.