SplitEasy: A Practical Approach for Training ML models on Mobile Devices

TL;DR

SplitEasy is a practical framework enabling mobile devices to train complex machine learning models using split learning, addressing key challenges and allowing minimal modifications for various neural networks, with efficient training times.

Contribution

This work introduces SplitEasy, a novel toolkit that facilitates split learning on mobile devices, overcoming theoretical and technical hurdles for practical deployment.

Findings

Successfully trained six neural networks on mobile devices using SplitEasy.

Achieved near-constant training time per data sample.

Enabled training of models that are otherwise infeasible on mobile devices.

Abstract

Modern mobile devices, although resourceful, cannot train state-of-the-art machine learning models without the assistance of servers, which require access to, potentially, privacy-sensitive user data. Split learning has recently emerged as a promising technique for training complex deep learning (DL) models on low-powered mobile devices. The core idea behind this technique is to train the sensitive layers of a DL model on mobile devices while offloading the computationally intensive layers to a server. Although a lot of works have already explored the effectiveness of split learning in simulated settings, a usable toolkit for this purpose does not exist. In this work, we highlight the theoretical and technical challenges that need to be resolved to develop a functional framework that trains ML models in mobile devices without transferring raw data to a server. Focusing on these challenges, we propose SplitEasy, a framework for training ML models on mobile devices using split learning. Using the abstraction provided by SplitEasy, developers can run various DL models under split learning setting by making minimal modifications. We provide a detailed explanation of SplitEasy and perform experiments with six state-of-the-art neural networks. We demonstrate how SplitEasy can train models that cannot be trained solely by a mobile device while incurring nearly constant time per data sample.