Trustworthy Federated Learning for Medical Imaging: A Structured Survey of Heterogeneity, Privacy Leakage, Inference Attacks, and Secure Collaboration

Abstract

Federated learning (FL) enables hospitals and imaging centers to train shared artificial-intelligence models while retaining raw patient data locally. This design is attractive for medical imaging because clinically useful models often require multi-institutional data, yet conventional pooling is constrained by privacy, governance, ownership, infrastructure, and regulatory requirements. Data locality, however, does not by itself establish trustworthiness. Federated systems can still fail under inter-hospital distribution shift, leak information through model updates or predictions, remain vulnerable to membership and reconstruction attacks, and incur substantial overhead when differential privacy or cryptographic defenses are added. This survey synthesizes the literature through five connected dimensions: predictive utility, heterogeneity robustness, formal privacy, empirical attack resistance, and system efficiency. It reviews foundations and architectures, application domains, non-IID optimization, privacy leakage, differential privacy, membership-inference auditing, secure aggregation, and cross-study evidence gaps. The central finding is that the field is comparatively mature within individual technical streams but weak at their intersection. Studies commonly demonstrate accuracy, non-IID robustness, privacy budgets, attacks, or cryptographic protection in isolation, while relatively few evaluate these properties under one consistent medical-imaging pipeline. The survey therefore proposes a compact five-dimensional taxonomy and a research agenda centered on patient-level integrity, realistic site heterogeneity, matched privacy baselines, strong empirical auditing, secure update aggregation, reproducible systems-cost measurement, and real multi-institutional validation.