Adaptive Differential Privacy Calibration for Federated Learning With Heterogeneous Client Data Distributions and Variable Participation Rates

This study investigates adaptive privacy budget calibration for federated learning with heterogeneous client data distributions and variable round participation rates under formal differential privacy guarantees within the context of privacy-preserving machine learning and distributed systems, an area of growing scientific importance given its implications for privacy-preserving healthcare AI, financial fraud detection, and cross-silo enterprise federated learning with regulatory compliance. Using adaptive per-round epsilon allocation via moments accountant tracking with client-heterogeneity-aware noise scaling and participation-rate-dependent gradient clipping, we examine heterogeneity-aware adaptive clipping norm reducing gradient distortion for underrepresented classes while maintaining formal epsilon-delta DP guarantees via moments accountant privacy composition in simulated federated settings with n=100-1000 clients, heterogeneity alpha=0.1-1.0 (Dirichlet), participation rate 10-30%, evaluated over 500 training rounds drawn from federated simulation framework with TensorFlow Federated and custom differential privacy accounting across MNIST, CIFAR-10, and chest X-ray classification tasks. Results indicate that adaptive calibration achieves 91.4% of non-private accuracy at epsilon=4 on CIFAR-10 with alpha=0.3 heterogeneity, a 14.2-point improvement over fixed-epsilon DP-FedAvg at equal privacy budget (p < 0.001), with 91.4% non-private accuracy at epsilon=4, 14.2 pp improvement over fixed-epsilon baseline as the primary quantitative benchmark. Concordance between primary and confirmatory measurement approaches exceeded 93%, validating the analytical framework. These findings contribute empirically to privacy-preserving machine learning and distributed systems and carry actionable implications for the design of programs and policies targeting privacy-preserving healthcare AI, financial fraud detection, and cross-silo enterprise federated learning with regulatory compliance.