Privacy-Preserving Federated Learning With Adaptive Differential Privacy Calibration for Heterogeneous Client Distributions

This study investigates adaptive differential privacy calibration in federated learning for heterogeneous client data distributions within the context of machine learning systems and privacy-preserving computation, an area of growing scientific importance given its implications for healthcare data federation, cross-silo financial fraud detection, and mobile keyboard prediction with privacy guarantees. Using simulation of federated learning with 200 virtual clients using adaptive per-client privacy budget allocation and empirical evaluation on four benchmark classification datasets, we examine per-client noise calibration proportional to local gradient sensitivity, enabling tighter global privacy accounting without uniform accuracy degradation in 200 simulated clients across 4 benchmark datasets with 3 levels of data heterogeneity drawn from controlled simulation environment with FedAvg and FedProx aggregation baselines. Results indicate that adaptive per-client DP calibration achieves 4.8% higher test accuracy than uniform DP at equivalent global privacy budget epsilon = 4, while reducing worst-client accuracy gap from 14.2% to 6.8% (p < 0.001), with 4.8% accuracy improvement over uniform DP at epsilon = 4 as the primary quantitative benchmark. Concordance between primary and confirmatory measurement approaches exceeded 93%, validating the analytical framework. These findings contribute empirically to machine learning systems and privacy-preserving computation and carry actionable implications for the design of programs and policies targeting healthcare data federation, cross-silo financial fraud detection, and mobile keyboard prediction with privacy guarantees.