A Federated Learning Framework For Personalized and Privacy-Preserving Biosignal Interfaces

Biosignal-controlled interfaces—using data collected from sensors in, on, or around the body such as cameras, accelerometers, or electromyography sensors—offer a richer way to serve individual abilities, preferences, and behaviors in mobile human-computer interactions. However, the centralized machine learning paradigm has several shortcomings which prevent us from realizing this vision. Centralized training demands that users share their sensitive biosignal data to train a one-size-fits-all model, which may lack accessibility and fail to generalize to novel or out-of-distribution users. Personalizing to improve performance via finetuning requires even more user data, further increasing the privacy risk. Personalized federated learning balances this privacy-personalization trade-off by leveraging information from peer users to enhance performance via personalized models without requiring users to upload their sensitive data. This work introduces a framework for researchers working with sensitive biosignal data to incorporate federated learning into their applications. We apply our framework to both an open-loop simulation and a closed-loop user study for an EMG interface. We show that traditional algorithms outperform our modified sequential FL algorithm in terms of performance, but have a substantially higher privacy risk. Finally, we highlight the need for more federated learning theory to support the sequential scenario inherent to many wearable applications.