Differential Privacy with Federated Learning via Entangled Noise

Abstract

Federated learning (FL) offers a promising paradigm for training machine learning models on decentralized datasets without directly exchanging the data itself. However, integrating differential privacy (DP) into FL introduces significant accuracy trade-offs. This work proposes a novel approach, "Entangled Noise," to mitigate these trade-offs. Entangled noise leverages a specialized noise injection strategy that minimizes communication overhead while maintaining a higher level of privacy guarantees. We demonstrate that standard DP methods often overly restrict model updates, leading to substantial accuracy degradation. Our entangled noise technique intelligently combines noise injections across model parameters, effectively reducing the overall noise variance and preserving model performance. We present a theoretical framework analyzing the privacy-utility trade-off and provide empirical evidence showcasing the superior performance of entangled noise compared to traditional DP methods in federated learning scenarios. The key innovation lies in the dynamic adjustment of noise parameters based on parameter sensitivity, resulting in a more efficient and accurate DP-FL system. This work contributes a practical and theoretically sound method for achieving stronger privacy guarantees in federated learning without sacrificing model accuracy.