Differential Privacy with Adaptive Noise Injection for Federated Learning

Abstract

Federated learning (FL) offers a promising approach to training machine learning models on decentralized data without directly exchanging the data itself. However, traditional differential privacy (DP) methods often introduce substantial noise into the model updates, significantly degrading model accuracy. This paper proposes a novel adaptive noise injection scheme for FL, designed to mitigate this issue. The core idea is to dynamically adjust the noise level based on the sensitivity of the local data, thereby optimizing the privacy-utility trade-off. Our approach leverages the inherent variance in local updates to reduce the overall noise required to achieve a desired privacy guarantee. We demonstrate through theoretical analysis and intuitive simulations that our adaptive scheme leads to improved model accuracy compared to standard DP methods while maintaining the same level of privacy. The proposed method addresses a critical bottleneck in FL—the inherent conflict between privacy preservation and model performance—and provides a practical pathway for deploying robust and accurate FL systems.