An On-device Federated Learning System for SMS Spam Classification

Federated Learning (FL) is a privacy-preserving machine learning technique that allows training of models on edge devices which are then aggregated centrally, ensuring private data never leaves edge devices. In this paper, we aim to bridge the research gap in on-device FL, by designing and testing an on-device FL system for classifying SMS spam on Android devices. We study the feasibility of training deep learning models on phones: optimizing models to fit within memory/network constraints, analyzing effects of different word embeddings, IID/non-IID data and finally optimizing FL performance to be on par with centralized training while preserving privacy.