Medical AI models trained on centralised patient datasets raise acute privacy concerns, creating tension between model accuracy and compliance with data protection regulations (HIPAA, GDPR).

To develop and evaluate a federated learning (FL) framework incorporating differential privacy (DP) for multi-class chest radiograph classification, assessing the accuracy-privacy trade-off across varying privacy budgets (epsilon).

A simulated distributed network of 8 hospital nodes each holding a local partition of a chest X-ray dataset (N = 12,000 images; 4 classes) was modelled. FL with Gaussian-mechanism DP (epsilon = 1, 5, 10, infinity) was implemented using PyTorch and Opacus.

At epsilon = 10, federated DP models achieved 91.4% classification accuracy (AUC = 0.943), only 3.1% below a centralised non-private baseline (94.5%). At epsilon = 1 (strongest privacy), accuracy dropped to 84.2%. Privacy budget epsilon = 5 offered the optimal accuracy-privacy balance.

FL with moderate differential privacy (epsilon = 5–10) provides an effective framework for privacy-preserving medical AI without unacceptable accuracy degradation, enabling cross-institutional model training without patient data centralisation.