Distributed Learning with Differential Privacy for Federated Edge Computing

Abstract

Federated learning (FL) offers a promising approach to training machine learning models on decentralized data sources, primarily located at edge devices. However, the inherent distributed nature of FL introduces significant challenges, particularly regarding privacy vulnerabilities and substantial communication overhead. This paper proposes a novel framework for distributed learning with differential privacy tailored for federated edge computing environments. The core idea is to integrate a customized differential privacy mechanism with a hierarchical learning structure and adaptive privacy budgets, all while exploiting hardware-accelerated cryptographic techniques to reduce communication costs. Our approach aims to simultaneously mitigate privacy risks and improve the efficiency of FL deployments across geographically diverse edge devices. We define the following key components: (1) A hierarchical learning structure consisting of local and global layers; (2) A novel differential privacy mechanism utilizing homomorphic encryption and secure multi-party computation; (3) Adaptive privacy budget allocation based on device heterogeneity and data sensitivity. The system is designed to minimize communication, maximize privacy, and improve the overall performance of FL in resource-constrained edge environments. The proposed method provides a theoretical framework for achieving strong privacy guarantees while maintaining reasonable communication costs.