Closed-loop Federated Learning for Continuous Forearm EMG-based Control

Invasive and non-invasive neural interfaces hold promise for rich, high-bandwidth inputs in next-generation technologies. However, these neural signals encode sensitive attributes about an individual's identity and health, and safely sharing this data for training neural decoders is a critical challenge.  Federated learning (FL) is a promising tool for privacy-preserving learning.However, FL's effects on privacy and performance for adaptive neural interfaces have not been explored.This work introduces FL for neural decoding and analyzes the effects of FL on open- and closed-loop decoder performance and data privacy using a high-dimensional electromyography interface. We find that in open-loop, FL outperformed a local-only baseline by 75\\%, demonstrating its promise for high performing, collaborative decoding.However, in our closed-loop user study, local decoders outperformed FL by 35\\% but with a 65\\% higher privacy risk, highlighting a practical performance-privacy tradeoff in closed-loop co-adaptation studies.These results offer empirical insights into the feasibility and limitations of FL for adaptive neural interfaces and point toward the need for new FL methods tailored to real-time, co-adaptive settings.