Federated evaluation: measuring model quality without sharing sensitive data
Technical framework for evaluating AI model performance across distributed data sources without centralizing sensitive information.
Authors
M. Santos, LEIBNIZ, 19 contributors
Published
2030
Citations
143
Overview
Healthcare systems, financial institutions, and government agencies often cannot share data for privacy/regulatory reasons, but need to evaluate model performance across these domains. This research develops evaluation frameworks that measure model quality without data sharing.
Methodology
Technical implementation across 6 healthcare networks and 4 financial institutions. Comparison of federated evaluation metrics against centralized evaluation using synthetic datasets. Performance benchmarking across evaluation methods. Security analysis and threat modeling.
Key Findings
Federated evaluation can measure most model quality metrics with 95%+ agreement to centralized evaluation, with the exception of tail-case performance (performance on rare but important cases). Tail-case performance requires either data sharing or dramatically larger federated cohorts (3-5x more participants).
Computational cost of federated evaluation is 4-6x higher than centralized evaluation, primarily in coordination overhead rather than computation. Privacy-computation tradeoff improves with more participants (with 50+ sites, overhead approaches 2x).
Five categories of evaluation metrics behave differently in federated settings: aggregate metrics (accuracy, AUC) work well; distribution-sensitive metrics (fairness metrics, performance across demographics) require careful design; fairness across distributed populations requires explicit weighting; calibration is difficult to verify in federated settings; rare-event detection requires sampling bias correction.
Implementing federated evaluation requires organizational alignment as much as technical work. Organizations accepting results from federated evaluation report higher trust than those relying on limited external validation, suggesting institutional legitimacy is achieved through inclusive evaluation processes.
Impact & Application
Deployed across 6 healthcare networks for model validation. Enables privacy-preserving model evaluation for sensitive domains. Used as standard practice in 12+ regulated industries.
Contributors
Lead: Dr. Marina Santos (Data Intelligence school). Collaborators from Mayo Clinic, Cleveland Clinic, JPMorgan Chase, and Bank of America. Technical advisors from Harvard Privacy Lab.