Momahmoses/federated-cancer-detection-africa

||Privacy-preserving federated cancer detection across 6 African hospitals where patient data never leaves the institution. # Pan-African Federated Cancer Detection Network A federated learning system that trains a shared breast cancer detection model across 6 hospitals in 6 countries, without a single patient record ever leaving any hospital. ## Problem 12 hospitals across 6 African countries want to train a shared AI model to detect breast cancer. But patient data cannot leave each country due to privacy laws (NDPR Nigeria, POPIA South Africa). Centralizing data is legally impossible. ## Quick Start ```bash pip install -r requirements.txt # Simulate 20 rounds of federated learning across 6 hospitals python run_federation.py ``` ## Architecture ``` Hospital A (Lagos) ┐ Hospital B (Nairobi) │ Local training only Hospital C (Cape Town) ├→ Weight updates (DP-noised) → Aggregation Server Hospital D (Accra) │ ↓ Hospital E (Kampala) │ Global Model (shared back) Hospital F (Addis) ┘ ``` **Patient data never leaves the hospital.** Only model weights travel. ## Components ### Federated Server (`src/federation/server.py`) - FedAvg: weighted average by number of local samples - Configurable: min clients per round, fraction participating - Flower (flwr) strategy or mock simulation mode - Tracks per-round loss and accuracy across all hospitals ### Hospital Client (`src/federation/client.py`) - Loads local patient data → trains → shares ONLY weights - Compatible with Flower's NumPyClient interface - Supports PyTorch or mock training ### Differential Privacy (`src/privacy/differential_privacy.py`) - **Gradient clipping**: bounds L2 norm of each update - **Gaussian noise**: calibrated to (ε, δ)-DP guarantee - **Secure aggregation**: server computes average without seeing individual updates - Default: ε=1.0, δ=1e-5 (strong privacy guarantee) ## Privacy Guarantee | Config | ε | Noise σ | Weight Error | |--------|---|---------|-------------| | High privacy | 0.5 | 2.16 | High | | Standard | 1.0 | 1.08 | Medi …