Diagnose with Confidence.
Protect with Certainty.
A secure, federated AI medical diagnosis platform built for Nigerian healthcare. Five hospital nodes train together. Zero patient data shared.
Supervisor: Dr. Ugo Nwankwo
How Federated Learning Works
Three steps. The model visits the data — the data never leaves the hospital.
Train Locally
Each institution trains the ResNet-50 model on their own chest X-rays. No image leaves the hospital.
Share Only Learning
Only model weight updates — protected by DP-SGD with gradient clipping (C=1.0) and Gaussian noise — are transmitted via mutually authenticated gRPC connections.
Improve Together
FedAvg aggregates updates weighted by dataset size. All five hospitals benefit from the improved model, which now reflects combined learning from 4,430 training images.
Mathematical, not just policy.
Three layered defences, each providing a verifiable guarantee.
Federated Learning
Patient data stays at source. The ResNet-50 model visits the data, not the other way around.
Differential Privacy
DP-SGD adds Gaussian noise N(0,σ²C²I) calibrated so the system satisfies (ε=5.13, δ=1e-5)-DP after 50 rounds. An adversary intercepting all model updates cannot determine any individual's participation to more than e^5.13 ≈ 169× baseline certainty.
Mutual TLS
All gRPC communications use mTLS 1.3 with X.509 certificates signed by the system's private CA. Certificate revocation via CRL.
Privacy-Utility Trade-off
AUC-ROC over 50 training rounds for each privacy configuration. Higher epsilon = weaker privacy guarantee. Moderate Privacy is the production default — 0.947 AUC-ROC for ε=5.13.
Ready to put privacy first?
Sign up as a patient for a symptom screening, or as a clinician to diagnose against the federated ResNet-50 model.