Federated Learning at the Edge
Multi-site NVFLARE deployment for privacy-preserving ML on heterogeneous hardware
Federated Learning at the Edge investigates the practical challenges of deploying privacy-preserving machine learning across multiple geographically distributed edge nodes — using NVIDIA FLARE (NVFLARE) on heterogeneous hardware typical of African research and healthcare institutions.
The Challenge
Centralised training requires pooling sensitive data (medical records, transaction logs, agricultural surveys) to a single server — untenable in many African contexts due to data sovereignty regulations, bandwidth constraints, and institutional trust barriers. Federated learning trains a shared model without ever moving raw data off-device, but real-world deployments face:
- Severe data heterogeneity (non-IID distributions across sites)
- Intermittent connectivity between nodes
- Compute heterogeneity (Jetson AGX vs. Raspberry Pi vs. server GPU in the same federation)
What We Built
- A 4-node NVFLARE federation spanning Raspberry Pi 4, Jetson Nano, Jetson Orin, and a desktop GPU server
- Custom FedProx and SCAFFOLD aggregation strategies tuned for high data heterogeneity
- Asynchronous round management to handle nodes dropping in and out of training rounds
- Benchmark suite comparing convergence speed, communication cost, and final model accuracy across aggregation strategies
Key Results
- FedProx with μ=0.01 converged 2.4× faster than FedAvg under the most heterogeneous data split
- End-to-end training on a simulated medical imaging task (chest X-ray classification) matched 94% of centralised accuracy with zero raw data sharing
- Asynchronous mode reduced total wall-clock time by 38% vs. synchronous rounds when one node had 3× slower compute