FedFarm: Privacy-Preserving Federated Learning for Cassava Disease Diagnosis on Smallholder Farmers' Smartphones

Abstract

Smallholder cassava farmers in sub-Saharan Africa face devastating crop losses from leaf diseases such as Cassava Mosaic Disease (CMD), Cassava Brown Streak Disease (CBSD), Cassava Bacterial Blight (CBB), and Cassava Green Mite (CGM), yet access to expert plant pathologists in rural areas remains critically limited. We present FedFarm, a Federated Learning (FL) system that empowers farmers to collaboratively train a cassava disease classifier directly on commodity Android smartphones, without ever transmitting raw images to a central server. This ensures the privacy of vulnerable agricultural communities while overcoming the bandwidth constraints typical of rural connectivity. FedFarm employs a split-model design in which a pretrained frozen MobileNetV2 backbone handles on-device feature extraction, while only a lightweight classification head (≃ 25 KB) is exchanged per communication round using the Federated Averaging algorithm, making the system viable even on 3G networks. In a real-world proof-of-concept deployment on two heterogeneous commodity devices, server-side classification accuracy improved from \(28.2\%\) to \(66.9\%\) across 100 rounds, well above the \(27.7\%\) majority-class baseline, while raw farmer data never left the device. FedFarm demonstrates that privacy-preserving, communication-efficient AI is achievable on the hardware that smallholder farmers already own, offering a replicable blueprint for responsible AI deployment in food-insecure, low-resource agricultural settings across the Global South.