Adaptive differential privacy in asynchronous federated learning for aerial-aided edge computing

Abstract

The integration of aerial-aided edge computing and federated learning (FL) is expected to completely change the way data is collected and utilized in edge computing scenarios, while effectively addressing the issues of data privacy protection and data distribution in this scenario. However, in the face of the challenge of device heterogeneity at the edge computing systems, most current synchronous federated learning approaches suffer from low efficiency because of the straggler effect. This issue can be significantly mitigated by adopting Asynchronous Federated Learning (AFL). Despite the potential benefits, AFL remains under-explored, posing a significant hurdle to optimizing the utility of privacy-enhanced AFL. To address this, we introduce adaptive differential privacy algorithms aimed at enhancing the balance between model utility and privacy in AFL. Our approach begins by defining two frameworks for privacy-enhanced AFL, taking into account various factors relevant to different adversary models. Through in-depth analysis of the model convergence in AFL, we demonstrate how differential privacy can be adaptively achieved while maintaining high utility. Extensive experiments on diverse training models and benchmark datasets showcase that our proposed algorithms outperform existing benchmark methods in terms of overall performance, enhancing test accuracy under similar privacy constraints and achieving faster convergence rates.