Federated Learning for Activity Recognition: A System Level Perspective

The past decade has seen substantial growth in the prevalence and capabilities of wearable devices. For instance, recent human activity recognition (HAR) research explored using devices applications such as remote monitoring patients, detection gait abnormalities, cognitive disease identification. However, data collection poses a major challenge developing HAR systems, especially because need to store at central location. This raises privacy concerns makes continuous difficult expensive due high cost transferring from user’s device repository. Considering this, we explore adoption federated learning (FL) potential solution address issues associated with HAR. More specifically, investigate performance behavioral differences between FL deep (DL) models, under various conditions relevant real-world deployments. Namely, two types models when (i) different sensor placements, (ii) having access users heterogeneous (iii) considering bandwidth efficiency, (iv) dealing incorrect labels. Our results show that suffer consistent deficit comparison their DL counterparts, but achieve these much better efficiency. Furthermore, observe exhibit very similar responses those exposed placements. Finally, are more robust labels than centralized counterparts.