Federated Learning in Open- and Closed-Loop EMG Decoding: A Privacy and Performance Perspective

Kai Malcolm, César Uribe, Momona YamagamiarXiv 2025 · DOI · arXiv

Abstract

Federated Learning (FL) is frequently proposed as a privacy-preserving solution for training neural decoders on sensitive biometric data like Electromyography (EMG). However, most evaluations rely on static, offline datasets. This paper presents the first empirical evaluation of FL in a real-time, closed-loop EMG decoding task. Our findings reveal a significant performance gap: while FL excels in offline simulations—improving tracking precision and privacy—it struggles in live settings where human-decoder co-adaptation occurs. We identify a "performance tension" in online user studies, suggesting that standard FL assumptions must be adapted for dynamic human-in-the-loop systems.

Introduction

High-dimensional neural signals, such as Electromyography (EMG), are essential for developing intuitive human-computer interfaces. However, these signals are also a "privacy nightmare," capable of revealing a user’s identity, age, sex, and medical conditions. This creates a significant privacy hurdle: while massive amounts of data are required to train robust decoders, users are understandably hesitant to share sensitive biometric data.

Federated Learning (FL) has emerged as a potential solution, allowing models to be trained across multiple users without moving raw data off local devices. While FL is often viewed as a "silver bullet" in offline settings, its efficacy in "closed-loop" scenarios—where a real human adapts to the decoder in real-time—remains largely unexplored. This study investigates whether FL can maintain its performance and privacy benefits when integrated into a live human-in-the-loop system.

Methodology

Learning Strategies

The researchers compared three primary learning strategies across both offline simulations and a live user study:

  1. Local Learning: The decoder is trained exclusively on the current user's data.
  2. Traditional FL (FedAvg): A single global model is trained by averaging updates from all participating users.
  3. Personalized FL (Per-FedAvg): A meta-learning approach designed to find a global initialization that can be rapidly fine-tuned for individual users.
Categorization of Federated Learning algorithms based on personalization and user availability.

Experimental Setup and Sequential FL

A key contribution of this work is the conceptualization of Sequential Federated Learning. Unlike standard lab settings where many users are connected simultaneously, real-world deployment often involves users appearing one by one. The authors modified Per-FedAvg to handle this sequential flow, where the global model acts as an evolving prior passed between participants.

Experimental setup for the 2D cursor trajectory tracking task using EMG sensors.

The study utilized two primary datasets:

  • Offline Dataset: A secondary analysis of data from 14 subjects performing a cursor-tracking task with a 64-channel EMG sensor.
  • Online User Study: A new real-time experiment with 16 participants (8 women, 8 men) performing the same task to observe closed-loop decoder adaptation.

Results

The Performance Gap

The study revealed a stark contrast between offline and online performance. In offline simulations, FL was a clear winner, improving tracking precision by up to 22% (approximately 0.232 cm/s better velocity error in cross-subject scenarios). However, in the real-time online study, the results flipped: local learning outperformed the modified Per-FedAvg by an average of 47.7 cm/s.

This discrepancy arises from the "Simulation Trap." In a live setting, the human and machine are co-adapting. The stabilizing effect of averaging across many users—a hallmark of FL—actually hindered the decoder's ability to stay synchronized with the shifting physiological patterns of a live user.

Privacy and Biometrics

The paper confirms that model weights themselves act as biometrics. An adversary could identify a user with 100% accuracy just by examining their trained local model weights. FL significantly mitigated this risk, reducing re-identification to ~33% for random initializations and ~25% for offline-pretrained initializations. This underscores FL's necessity for privacy, even if it introduces a performance cost in online settings.

Discussion and Future Directions

The erratic fluctuations observed in online decoders, compared to the smooth exponential decay of offline convergence, highlight the noise inherent in real-time interaction. Furthermore, the authors note that moving to complex, non-linear models like deep neural networks would likely exacerbate privacy risks, as these models have a higher capacity to memorize unique biometric fingerprints.

Future research must move beyond "one-size-fits-all" FL algorithms. There is a critical need for adaptive hyperparameter strategies that can dynamically balance global stability for privacy with the high degree of personalization required for stable, real-time neural control. The field must develop FL frameworks specifically designed for the unique dynamics of human-in-the-loop systems.

Summary updated

BibTeX
@misc{malcolm2025federated,
  author = {Kai Malcolm and César Uribe and Momona Yamagami},
  title = {{Federated Learning in Open- and Closed-Loop EMG Decoding: A Privacy and Performance Perspective}},
  year = {2025},
  eprint = {2507.12652},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.12652},
}
Figure