TinyFed: Lightweight Federated Learning for Constrained IoT Devices

Tiny Machine Learning (TinyML) is a paradigm that directly enables the inference of machine learning (ML) algorithms on microcontrollers — devices with limited computational resources, including energy, processing power, and storage constraints. Extending TinyML to allow inference and model training in a collaborative setting has led to the development of Tiny Federated Learning (TinyFL). This proposal introduces TinyFed, a lightweight framework designed to support the entire Federated Learning cycle, including training a neural network on the device, transmitting the model to a server, aggregating multiple models, and delivering the aggregated model back to resource-constrained microcontrollers engaged in the federated process. TinyFed has been validated by deploying a neural network with four inputs, three hidden layers (16, 8, and 4 neurons), and two outputs on ESP32 devices to detect temperature, humidity, luminosity, and voltage measurement anomalies. Our results demonstrate the feasibility of this approach with successful local training.