Synergistic differential privacy-enhanced federated learning framework for heterogeneous iot environments

Abstract

Collaboratively training machine learning models on resource-constrained edge devices, while ensuring user data privacy, has become a key and challenging research challenge. Federated learning (FL) provides a basic framework for this purpose, however, its practical application still faces complex, interconnected challenges due to the introduction of differ-ential privacy (DP) noise, imperfectly distributed data (Non-IID), and com-munication bottlenecks. Most existing research focuses on solving one of these problems in isolation, often ignoring the complex trade-offs between them.