Semantic-Aware Differentially Private and Trust-Driven Robust Federated Optimization for Multi-Center Medical Image Analysis

Abstract

Federated learning enables hospitals to jointly train diagnostic models without sharing raw medical images, yet multi-center deployment remains vulnerable to site-dependent semantic shifts, privacy leakage, and unreliable model updates. This paper presents a two-level optimization framework. At the client level, SemaDiffPriv disentangles shared diagnostic representations from institution-specific features and perturbs only clipped shared updates, retaining private semantics locally under differential privacy. At the server level, FedTrustOpt estimates client reliability from update stability, historical consistency, and directional agreement, producing trust scores that control both robust aggregation weight and the next-round privacy budget. A result-free evaluation protocol is specified for multi-center brain-tumor segmentation under non-IID data and anomalous-client settings, with formal definitions of convergence variance, angular update dispersion, and robustness degradation. The framework establishes a unified basis for studying utility, privacy, and aggregation robustness without centralized medical data.