Adaptive Privacy Budget Allocation Optimization for Multi-Institutional Federated Learning in Healthcare

Abstract

Multi-institutional healthcare collaborations increasingly rely on federated learning for privacy-preserving machine learning across distributed medical datasets. Traditional uniform privacy budget allocation strategies fail to account for data heterogeneity among institutions, leading to suboptimal model utility. This paper proposes an Adaptive Privacy Budget Allocation algorithm that dynamically distributes privacy budgets based on quantifiable institutional data heterogeneity measures. The Data Heterogeneity Index captures variations in label distributions, feature characteristics, and dataset sizes. Experimental evaluations on three medical datasets demonstrate accuracy improvements of 3.2-5.8 percentage points and privacy budget efficiency gains of 28-41% compared to baselines. The framework provides practical guidance for healthcare institutions balancing privacy protection with collaborative learning effectiveness.