FedFarm is a mobile-oriented Federated Learning (FL) system designed to classify cassava leaf diseases directly on Android devices. It represents a comprehensive FL solution tailored to this specific agricultural challenge. The system operates on standard Android smartphones, utilizing a frozen MobileNetV2 backbone running via PyTorch Mobile as a feature extractor—eliminating the need for on-device training of the base network.
Instead, it collaboratively trains a lightweight linear classification head across the fleet of devices using the Federated Averaging (FedAvg) algorithm, coordinated by the Flower FL framework via gRPC transport. This architecture enables distributed model training over a network of devices in the field without centralizing raw image data, thereby preserving farmer privacy while continuously improving a shared disease detection model.