Transfer Learning for Human Activity Recognition in Federated Learning on Android Smartphones with Highly Imbalanced Datasets

Federated Learning (FL) on edge devices optimizes network resources while addressing privacy concerns. However, the model complexity is a challenge. Nonindependent and identically distributed (non-IID) data presents extra challenges for model convergence. Furthermore, Human Activity Recognition (HAR) is gaining traction thanks to advancements in sensors and AI. The present work introduces a method using transfer learning for FL on Android smartphones with highly imbalanced HAR datasets, in two phases. First, a model is trained offline using a large amount of data. This base model serves as a feature extractor for the head model, which is trained with local data on the FL clients. The prior knowledge acquired by the base model allows a tiny head model, leading to small latencies during the FL phase while maintaining good accuracy. The approach's effectiveness is demonstrated through an experiment on an FL setup involving twelve mobiles in five cross-validation folds. Profiling metrics during training for each mobile type are presented.