Enhanced Shuffle Models for Optimized Differential Privacy in Federated Learning

The core innovation is a refined shuffle model using a secure “invisible cloak”-based protocol. This eliminates the need for a trusted shuffler and simplifies data security without complex cryptography. IS-FL also features a new mechanism for random selection and noise injection during training. By selectively applying Laplacian noise to gradient data, it safeguards strong privacy while minimizing accuracy loss. Experiments on real-world datasets show that IS-FL outperforms traditional DP-FL, LDP-FL, and the state-of-the-art SS-FL. At the same privacy budget, IS-FL has notably higher test accuracy. For example, at a privacy budget of (2.0, 5 e-6), it retains 99% of non-privacy-preserving FL accuracy, showing its excellent privacy-performance balance.