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

1. Problem Statement

Neural interfaces, particularly those using high-dimensional electromyography (EMG) signals, offer high-bandwidth control but face a critical tension between data needs for training and user privacy.

• Privacy Risks: Neural signals contain unique physiological signatures (biometric markers) that can reveal sensitive attributes like identity, age, sex, and even typed passwords (acting as a "biological keylogger"). Centralizing this data for training creates significant re-identification risks.
• The Gap: While Federated Learning (FL) is a promising privacy-preserving solution that keeps raw data local, its efficacy in real-time, online neural interfaces remains unexplored. Most FL research relies on offline simulations with static datasets.
• The Challenge: Online neural interfaces involve human-decoder co-adaptation, where the user and the decoder continuously influence each other. Standard FL assumptions (e.g., synchronous user availability, static data distributions) may fail in this dynamic, sequential environment, potentially destabilizing the optimization process.

2. Methodology

A. Conceptual Framework

The authors propose a framework to guide FL algorithm selection based on two key attributes:

1. Personalized Cost: Whether the data is heterogeneous enough to require Personalized Federated Learning (PFL) rather than Traditional FL (TFL). Neural signals are highly heterogeneous, suggesting PFL is necessary.
2. User Availability: The mode of user interaction. The authors define three modes:
   • Synchronous: All users available simultaneously (standard FL).
   • Asynchronous: Users available at different times (common in deployment).
   • Sequential: A novel contribution. Only one user is active at a time, and they appear exactly once (typical in lab-based online studies). This breaks the core FL assumption of aggregating updates from multiple users simultaneously.

B. Experimental Design

The study evaluates EMG decoding across two settings using a cursor control task (tracking a reference trajectory with forearm muscles):

1. Offline Study (Simulation):

   • Data: Secondary analysis of data from 14 subjects (Madduri et al.).
   • Algorithms: Compared Local (non-federated), FedAvg (Traditional FL), and Per-FedAvg (Personalized FL).
   • Setup: Simulated synchronous training with k-fold cross-validation (Intra-subject and Cross-subject scenarios).

2. Online Study (Real-time):

   • Participants: 16 new subjects.
   • Algorithms: Compared Local, Modified Per-FedAvg (adapted for sequential use), and Static (no adaptation).
   • Adaptation: Since no FL algorithms exist for the "Sequential" scenario, the authors modified Per-FedAvg. Instead of averaging multiple users, the global model is updated via a weighted average between the single active user's weights and the server's current global weights (mathematically equivalent to SmoothBatch regularization).
   • Initialization: Tested with both Random initialization and initialization from the offline Per-FedAvg global model.

C. Evaluation Metrics

1. Velocity Error: The primary performance metric, measuring the deviation between the predicted cursor velocity and the optimal velocity required to track the target.
2. Model Re-identification Risk: A privacy metric quantifying an adversary's ability to identify a user's identity solely from the trained decoder weights (using a Support Vector Classifier).

3. Key Contributions

1. First Empirical Evaluation of Online FL: The study provides the first real-time user study comparing FL against local learning in neural decoding, moving beyond offline simulations.
2. Definition of Sequential FL: The authors identify and define "Sequential Federated Learning" as a distinct regime where users train one-by-one, highlighting that standard averaging-based aggregation is impossible without violating privacy or storing raw data.
3. Privacy-Performance Trade-off Analysis: The study reveals that the relationship between privacy and performance differs drastically between offline and online settings.
4. Conceptual Framework: A guide for researchers to select FL algorithms based on data heterogeneity and user availability constraints.

4. Key Results

Offline Results (Simulation)

• Performance: Both FL algorithms (FedAvg and Per-FedAvg) outperformed Local learning, reducing velocity error by up to 22%.
• Privacy: FL achieved a 0.0 re-identification risk (perfect privacy), whereas Local learning had a 1.0 risk (fully compromised).
• Convergence: FL decoders converged smoothly and consistently across users.

Online Results (Real-time)

• Performance Reversal: Local learning significantly outperformed the modified Per-FedAvg algorithm. Local decoders reduced velocity error by approximately 43.7 cm/s compared to FL.
• Privacy: Despite lower performance, the FL approach still offered better privacy (lower re-identification risk) than Local learning.
• Convergence: Online FL decoders exhibited fluctuations and failed to converge to a stable solution, often staying close to their initialization. Local decoders showed dispersed trajectories, indicating successful user-specific adaptation.
• Initialization: The initialization method (Random vs. Offline) did not significantly impact performance, suggesting the online dynamics override the benefits of pre-training.

5. Significance and Discussion

• The "Offline-to-Online" Gap: The study demonstrates that results from offline FL simulations do not generalize to real-time neural interfaces. The co-adaptive nature of online decoding (where the user changes their motor strategy based on decoder performance) creates a dynamic environment that standard FL hyperparameters cannot handle.
• Limitations of Current FL: Standard FL relies on inter-user regularization (averaging updates from many users) to stabilize training. In sequential online settings, this mechanism is lost, leading to instability and poor performance when forcing a global model onto a single user's real-time adaptation.
• Privacy vs. Utility: While FL successfully protects privacy in online settings, it currently incurs a heavy performance penalty. This creates a "performance-privacy trade-off" that is not present in offline simulations.
• Future Directions: The paper argues for the development of specialized adaptive federated algorithms designed specifically for sequential, real-time, co-adaptive environments. These algorithms must balance global regularization with the high degree of personalization required for stable neural control.

Conclusion

Federated Learning is a viable path for privacy-preserving neural interfaces, but standard FL methodologies are insufficient for real-time online decoding. The study concludes that while FL offers superior privacy, it currently struggles to match the performance of local, co-adaptive learning in live human-in-the-loop scenarios. Future work must focus on developing algorithms that can navigate the unique dynamics of sequential-user neural decoding without sacrificing the stability required for real-time control.