DynamicNet: Efficient Federated Learning for Mobile Edge Computing With Dynamic Privacy Budget and Aggregation Weights

Mobile Edge Computing (MEC) has been successfully applied to smart homes, mobile devices, and other electronic products, offering reduced communication latency and alleviating pressure on cloud services. The integration of Federated Learning (FL) and MEC provides customized services for enhancing user experiences. However, there are still gaps in combining these technologies: FL is susceptible to membership inference attacks, risking user privacy; MEC faces limitations in energy and computational resources, restricting FL performance. In order to address these issues, we propose a novel approach, named DynamicNet, which incorporates Privacy Budget Regulation (PBR) and Dynamic Aggregation Weights (DAW). PBR addresses privacy leaks in edge devices by assessing local model privacy, dynamically adjusting privacy budget for subsequent use, ensuring desired level of privacy protection while minimizing noise impact on global model. DAW tackles MEC-related performance limitations by generating synthetic data, computing optimal weights for edge devices using this synthetic data, and aggregating parameters accordingly. Experimental results on two mainstream datasets demonstrate: 1) PBR preserves participant data privacy while enhancing global model convergence speed and accuracy; 2) DAW adapts well to heterogeneous federated environments, ensuring high global model accuracy even when some edge device datasets are of inferior quality.