Inclusive federated learning in medical AI via compliance-weighted noise allocation

The researchers propose a compliance-weighted federated learning framework that adjusts server-side DP-SGD noise according to institutional compliance scores, allowing low-compliance institutions to participate in medical AI training without dragging down other parties. On PneumoniaMNIST and BreastMNIST (16 clients), under equal-mean noise it achieves utility comparable to uniform DP; after including 12 low-compliance institutions, BreastMNIST pooled accuracy +2.8 pp (n=5, not significant). The formal guarantee covers only the aggregator dataset; client-level DP requires secure aggregation.