Privacy-aware collaborative machine learning for fitness tracking: A New Federated Learning Baseline

What is it about?

Today, many wearable devices such as smartwatches and rings use built-in sensors to track our physical activities for health and fitness. Usually, these devices send our private movement data to a central server to help "train" the applications to recognise different exercises. However, this method raises significant privacy and security concerns for users. To solve this, we developed FedFitTech, a new system that uses a method called Federated Learning. This technology allows your fitness tracker to learn from your movements without ever sending your raw, private data away from your personal device. Instead, the device processes the information locally and only shares small "intelligence updates" about what it has learned with a central system. This approach keeps sensitive information safe while still allowing the app to get smarter by combining patterns from many different users. Our research also introduces a special "early stopping" feature. This allows the device to stop communicating with the server once it has learned enough, saving battery life and reducing data usage by approximately 13% while keeping the app’s accuracy nearly the same. FedFitTech is now available as an open-source tool to help other researchers and developers create more private, secure, and efficient fitness technology for everyone.

Featured Image

Photo by Vitaly Gariev on Unsplash

Photo by Vitaly Gariev on Unsplash

Why is it important?

This work addresses the growing tension between the widespread adoption of wearable fitness technology and the increasing demand for data privacy. While traditional Centralized Learning methods struggle with regulatory restrictions like GDPR and CCPA, this paper introduces FedFitTech, a specialised Federated Learning baseline that allows models to be trained on-device without sharing sensitive raw sensor data.