TinyFed: Lightweight Federated Learning for Constrained IoT Devices

Abstract

TinyML enables machine learning inference on microcontrollers with limited resources. Extending this to a collaborative setting led to Tiny Federated Learning (TinyFL). This article presents TinyFed, a lightweight framework that supports the full federated learning cycle—from local training to model aggregation and redistribution. TinyFed was validated on ESP32 devices using a neural network with four inputs, three hidden layers, and two outputs to detect temperature, humidity, luminosity, and voltage anomalies. Local training achieved accuracies of up to 99.47%.