TinyFedTL: Federated Transfer Learning on Ubiquitous Tiny IoT Devices

Abstract

TinyML refers to the intersection of machine learning (ML), mathematical optimization, and tiny IoT embedded systems. In the current era of ubiquitous connectivity and pervasive data, TinyML has emerged as an effective method to continuously analyze real-world data without the resource overhead of traditional ML hardware. However, as valuable data around us is subjected to strict privacy and security guarantees, there is a distinct lack of ML solutions that operate under both constrained computing environments and privacy concerns.In light of these challenges, we present TinyFedTL1, the first open-sourced implementation of federated learning (FL) at the most resource-constrained level in IoT containing microcontroller unit (MCU) and small CPU based devices. Using transfer learning (TL) as a representative task, we demonstrate that privacy-centric FL on devices with a tiny memory footprint (less than 1MB) is not only possible but also effective. Researchers and engineers can use TinyFedTL to open up data across various fields that can be used to gain insights for improving quality of life or user experience without sacrificing privacy.