AI
Federated Healthcare AI Lets Lower-Compliance Institutions Join Without Sacrificing Accuracy
Researchers introduce a compliance-aware privacy approach that adjusts noise levels per institution, allowing less-regulated sites to participate in collaborative model training without uniform penalties.
Federated learning promises collaborative clinical AI development while keeping patient data on-site, but a key barrier remains: how to balance privacy protection across institutions with vastly different compliance resources and capabilities.
A new framework addresses this tension by tailoring privacy noise to each institution's compliance posture, rather than applying a one-size-fits-all approach. Researchers at arXiv cs.AI developed a compliance-weighted noise allocation system that maps institutional compliance levels—measured against HIPAA, GDPR, NIST, ISO, and HL7/FHIR standards—to per-institution privacy noise scales.
How the Approach Works#
The method assigns each participating site a compliance score, which then determines the amount of Gaussian noise applied during federated learning training. Better-compliant institutions receive less noise, while lower-compliance sites receive higher noise to compensate for their weaker controls. This targeted allocation operates on the aggregator dataset, the centralized system that combines updates from all sites.
The researchers tested five standard federated strategies—FedAvg, FedMedian, FedProx, FedYogi, and FedAdam—on medical imaging datasets (PneumoniaMNIST and BreastMNIST) across 16 virtual clients over 50 training rounds.
Results: Inclusion Without Trade-offs#
The headline finding: including 12 lower-compliance sites in the training pool, versus restricting participation to compliant-only institutions, boosted accuracy on BreastMNIST by an average of +2.8 percentage points across strategies (with gains ranging up to +17 pp in certain configurations). The largest improvements came from FedMedian (+6.8 pp) and FedProx (+5.2 pp), though FedAdam showed a slight decline.
Critically, the compliance-weighted noise approach matched the performance of uniform server-side privacy at equal mean noise levels, incurring no additional accuracy penalty. The initial noise cost in the first training round was modest—1.3 percentage points for BreastMNIST and 2.5 pp for PneumoniaMNIST.
Privacy Guarantees and Future Work#
The framework provides formal differential privacy bounds (ε,δ) for the aggregator dataset. Cumulative privacy budgets reached approximately 1434 for BreastMNIST and 513 for PneumoniaMNIST at δ=10⁻⁵. However, the current approach does not yet guarantee client-level privacy—protecting individual institutions' local data—which would require secure aggregation protocols in future work.
