Paper Title
EXPERIMENTAL EVALUATION OF ADAPTIVE PRIVACY BUDGET SCHEDULING IN FEDERATED LEARNING FOR MEDICAL IMAGE ANALYSIS

Abstract
Federated Learning (FL) allows collaborative training of models among multiple medical institutions while keeping the sensitive patient data decentralized. However, incorporating Differential Privacy (DP) into FL introduces a trade-off between model utility and privacy protection. Most existing DP-based FL systems utilize static privacy budgets, using a constant privacy parameter (ε) across all communication rounds and clients. Such an inflexible approach fails to account for the changing dynamics of training, which can result in inefficient privacy utilization. In our prior work we proposed “Dynamic Privacy Budget Allocation Framework (DPBAF)” that utilizes adaptive privacy scheduling as a constrained optimization problem under Rényi Differential Privacy (RDP) accounting. While the previous work established the theoretical foundation, it did not provide empirical validation. In this paper we explore experimental extension of DPBAF within a controlled federated environment on medical images. We evaluate the framework on BloodMNIST and PathMNIST datasets using 5 clients over 40 communication rounds. The experiments are performed using 4 different seeds and comparison is made between static (ε) and adaptive (ε). Results averaged across seeds demonstrate that DPBAF consistently improves final model accuracy while maintaining bounded cumulative privacy loss under RDP composition. The results support the theoretical claim that adaptive privacy redistribution enhances privacy-utility efficiency without significantly compromising privacy constraints. Keywords - Federated Learning, Differential Privacy, Dynamic Privacy Budget Allocation, Rényi Differential Privacy, Medical Image Classification.