Dynamic Optimization Method for Differential Privacy Parameters Based on Data Sensitivity in Federated Learning

Abstract

This paper introduces a dynamic optimization framework for differential privacy parameters in federated learning systems that adapts privacy budgets based on real-time data sensitivity assessment. The proposed methodology employs a lightweight sensitivity analyzer that categorizes data samples into predefined tiers through statistical and semantic feature extraction, enabling granular privacy budget distribution across heterogeneous clients. An adaptive noise calibration algorithm dynamically modulates Gaussian noise injection based on sensitivity assessments and model convergence metrics. Experimental validation across four datasets from healthcare and financial domains with 120 federated clients demonstrates that our approach achieves 96.1% accuracy at ε=1.9, outperforming a static differential privacy baseline at the same ε (73.2%), representing a (relative +31.3%, +22.9pp) improvement in model utility. The framework reduces privacy budget exhaustion by 45% and extends training duration by 69% while maintaining formal differential privacy guarantees. Performance analysis shows the method maintains robustness under non-IID data distributions with an alpha = 0.1 Dirichlet parameter and tolerates up to 20% malicious clients through Byzantine-robust aggregation.