FedLRDP: Federated Learning Framework with Local Random Differential Privacy

Federated learning (FL) is a distributed machine learning framework enabling multiple clients to collaboratively train a shared ML model without sharing raw data. Despite its aim to safeguard data security and privacy, FL faces risks from advanced adversarial attacks like membership inference attack (MIA), potentially leaking sensitive information. To counter these threats, differential privacy (DP) methods add noise to shared model parameters. However, traditional DP methods struggle with balancing model accuracy, training efficiency, and privacy protection, often compromising one for the other. Additionally, uniform DP mechanisms may not cater to varying client privacy and performance needs. To tackle these challenges, we propose a federated learning framework based on localized random differential privacy (FedLRDP). This approach empowers each client to control noise levels based on their privacy requirements, enhancing model performance while preserving data privacy. We further optimize client-side loss functions to enhance model performance. Theoretical analysis establishes the convergence bounds for FedLRDP, demonstrating its superior convergence performance. Experimental results on MNIST reveal that FedLRDP improves model accuracy by 3.28% compared to traditional DP methods while mitigating MIA with a 72.48% lower attack success rate than FedAvg. Moreover, on EMNIST, FedLRDP boosts model accuracy by 7.31% compared to traditional DP methods, maintaining a 59.4% lower attack success rate than FedAvg. These findings underscore FedLRDP’s efficacy in meeting client privacy needs while enhancing model performance.