This work is licensed under a Creative Commons Attribution-NonCommercial International 4.0 License. PETRA ’24, June 26–28, 2024, Crete, Greece ˝ 2024 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-1760-4/24/06 https://doi.org/10.1145/3652037.3652074 ABSTRACT The privacy of personal data is paramount in the realm of assisted living and digital healthcare. Federated Learning (FL), with its decentralised model training approach, has emerged as a compelling solution to reconcile the need for personalised models with the requirement to protect sensitive personal information. By allowing model training to occur locally on user devices without centralising raw data, FL is intended to strike a balance between personalisation and privacy. While the potential benefits of FL in assisted living and digital healthcare are substantial, practical implementation poses significant challenges. One of them is the non-Independently and Identically Distributed (non-IID) nature of personal data. Unlike centralised datasets, non-IID data exhibits inherent variability across different individuals, as well as their surrounding contexts. Unfortunately, many research approaches in this domain often overlook the nuances of non-IID data, potentially leading to models that lack robust generalisation across diverse healthcare scenarios. To highlight the importance of this challenge, in this paper, we report on our hands-on experience of building a FL system for drowsiness detection using non-IID data.We compare this federated setup with a traditional, centralised approach to model training by identifying and discussing the associated challenges from multiple perspectives, as well as possible solutions and recommendations for further research. AUTHORS Rustem Dautov rustem.dautov@sintef.no SINTEF Digital Oslo, Norway Erik Johannes Husom erik.johannes.husom@sintef.no SINTEF Digital Oslo, Norway