Federated Learning in Offline and Online EMG Decoding: A Privacy and Performance Perspective

TL;DR

This study proposes a conceptual framework for applying FL to the distinct constraints of neural interface application and provides a systematic evaluation of FL-based neural decoding using high-dimensional electromyography (EMG) across both offline simulations and a real-time, online user study.

Abstract

Neural interfaces offer a pathway to intuitive, high-bandwidth interaction, but the sensitive nature of neural data creates significant privacy hurdles for large-scale model training. Federated learning (FL) has emerged as a promising privacy-preserving solution, yet its efficacy in real-time, online neural interfaces remains unexplored. In this study, we 1) propose a conceptual framework for applying FL to the distinct constraints of neural interface application and 2) provide a systematic evaluation of FL-based neural decoding using high-dimensional electromyography (EMG) across both offline simulations and a real-time, online user study. While offline results suggest that FL can simultaneously enhance performance and privacy, our online experiments reveal a more complex landscape. We found that standard FL assumptions struggle to translate to real-time, sequential interactions with human-decoder co-adaptation. Our results show that while FL retains privacy advantages, it introduces performance tensions not predicted by offline simulations. These findings identify a critical gap in current FL methodologies and highlight the need for specialized algorithms designed to navigate the unique co-adaptive dynamics of sequential-user neural decoding.

Figures (5)

• Figure 1: Experimental setup. The subject controls a blue cursor using their forearm muscles, with activity recorded from a 64-channel Quattrocento EMG sensor. The subject's goal is to track the red target by minimizing the distance between the blue cursor and the red reference trajectory.
• Figure 2: Offline data splits: Local train (blue), FL train (grey), and test (orange) split for the first fold in seven-fold cross-validation ($N=14$ subjects). For the Intra-Subject Scenario, the federated decoder is trained across six training folds for all 14 subjects. Each subject-specific Local decoder is trained on a single subject's six training folds and evaluated on that subject's withheld testing fold. For the Cross-Subject Scenario, the federated decoder is trained on the entire dataset for Subjects 1-12, and cross-subject decoders are evaluated on Subjects 13-14. Likewise, the Local decoders are trained on a single subject's data, and all are tested on the same withheld subjects.
• Figure 3: Offline performance (top) and privacy risk (bottom) for Local, and Per-FedAvg Algorithms ($N=14$ participants). Scenarios separated: Intra-subject (left) and Cross-subject (right). For all plots, dots indicate the median and error bars represent the interquartile range (IQR: 25th, 75th percentile). For both metrics, lower values indicate better performance and improved privacy risk.
• Figure 4: Experiment performance (Top) and privacy risk (bottom) for Local, Per-FedAvg, and Static Algorithms with Random (light blue), offline (dark blue), and Pretrained (green) decoder Initializations ($N=16$ participants). For all plots, dots indicate the median and error bars represent the interquartile range (IQR: 25th, 75th percentile). For both metrics, lower values indicate better performance and improved privacy risk.
• Figure 5: Top: Euclidean distance from the given update's decoder to the final decoder, for both offline (left) and online (right) evaluations. For the purposes of visualization and comparison, we solely plotted the Per-FedAvg (PFA) algorithm for both offline (Intra-subject scenario) and online (Random Initialization) evaluations. Bottom: PCA applied to each update's decoder for all trials and plotted in a two-dimensional sub-space. Initial decoders are denoted with a circle marker, final decoders are denoted with x markers. The average final decoder, averaged across all users within each condition, is denoted with a star marker. For the purposes of clear visualization, only the Intra-subject scenario results are plotted for offline, and only Random Initialization results are plotted for online. The corresponding results (i.e., Cross-subject scenario results for offline, offline Initialization for online) followed nearly identical trajectories. The full plot can be found in the Supplementary Materials.