Wearable decoders for muscle signals (electromyography, or EMG) depend heavily on where the electrodes sit. Moving a high-density array 1 cm can cut gesture accuracy by up to 25% (Tackling Electrode Shift in Gesture Recognition with HD-EMG Electrode Subsets). Federated learning trains one shared model across many users' devices by averaging their weight updates, so raw recordings never leave the device. It quietly assumes that "channel 17" means the same patch of skin on every device. When each user's sleeve sits a little differently, that assumption fails. Averaging then blends weights that were tuned to different muscles, and part of what looks like irreducible between-user variation is really a misaligned coordinate system.
We propose to give each device a tiny private registration: two or three numbers, such as how far along the forearm and how far around it the array sits. These numbers resample the local recording into a shared reference frame before it reaches the federated model. The registration is estimated on the device and never transmitted. Only the decoder that operates in the reference frame is averaged. In plain terms, every device lines up its own signals before collaborating, and it does not have to reveal how it lined them up.
Core Claim
Our claim is that, in federated EMG decoding, a few private per-device placement offsets learned jointly with a shared reference-frame decoder will outperform two alternatives. The first is making the shared model invariant to placement through augmentation. The second is giving each device a fully private input layer. We expect the margin to be largest under rotation of the array around the limb. Rotation moves the array onto different muscle groups, so a model that is invariant to it must discard exactly the spatial information that separates gestures. A model that registers the shift keeps that information.
The objective is a nested problem:
$$\min_{\theta}\ \sum_{k=1}^{K} \frac{n_k}{n}\ \min_{\phi_k}\ \Big[\mathcal{L}_k\big(f_\theta \circ R_{\phi_k}\big) + \lambda \lVert \phi_k \rVert^2\Big].$$
Here $f_\theta$ is the decoder shared across devices, with parameters $\theta$. $R_{\phi_k}$ resamples device $k$'s electrode grid into the reference grid using its private offset $\phi_k$, which holds a longitudinal shift and a circumferential angle. Rotation is treated as cyclic where the array wraps the limb and masked where coverage ends. $\mathcal{L}_k$ is the device's local task loss, $n_k$ is its sample count, and $\lambda$ is a small prior pulling offsets toward zero. The prior also fixes the reference frame, which is otherwise defined only up to one global shift.
Relation to Prior Work
Three centralized lines of work motivate the registration idea, and each also constrains it:
- Calibrated ECG field model. NEF-NET+ (Nef-Net v2) fits learnable angular deviations from the first 5 seconds of a misplaced ECG to register leads to a continuous cardiac field. This is the closest mechanism to ours. It relies on CT-derived coordinates and fixed limb anchor leads, and it is trained centrally.
- Coordinate-aware transformer. The Population Transformer (Population Transformer) adds known 3D electrode coordinates to per-channel embeddings so that arbitrary layouts can be pooled. It too assumes the coordinates are known and the data are centralized.
- Subset augmentation. All Valid Subsets (Pereira et al.) takes the opposite route and buys invariance by training on shifted sub-grids. It handles only proximal–distal shifts.
Our method borrows the "estimate the offset, then resample" step from NEF-NET+. It differs in three ways: no coordinates are known, the offset is learned through the task loss, and the offset is kept private.
On the federated side, two papers constrain the design:
- Federated EMG. Malcolm et al. run FedAvg, Per-FedAvg (a meta-learned starting point for fast local adaptation) and sequential federated learning on a fixed 64-channel forearm grid. They show that locally trained weights re-identify users with 100% accuracy, and that averaging reduces this to roughly 25–33%. Shared geometry is therefore a privacy liability, which is why our offsets stay on the device. Malcolm et al. never move the electrodes.
- Federated EEG. mixEEG improves federated cross-subject EEG through channel and frequency mixup. It still treats channels as abstract indices on a matched cap.
Neither paper separates placement from physiology. Separating them is what we propose to do.
Research Program
We will build the resampling operator $R_\phi$ as differentiable interpolation on each dataset's physical grid, cyclic in the circumferential direction where the array wraps the limb. Training alternates two steps:
- Each device fixes $\theta$ and updates $\phi_k$ by a few gradient steps on its local data.
- It then fixes $\phi_k$, updates $\theta$, and sends only the change in $\theta$.
A new user arrives with no offset. It estimates $\phi$ from a short calibration recording, either a few seconds of labeled gestures or unlabeled data with entropy minimization. Only after that does any fine-tuning happen.
The first scientific question is whether placement accounts for a measurable share of between-device heterogeneity. We will test this by inducing known shifts on public data and measuring how much FedAvg degrades beyond the subject-only case.
The second question is whether the learned offsets recover the induced shifts. Because we inject the shifts, the true offset is known, so registration error can be reported in millimetres and degrees.
The third question concerns privacy. The decoder sees resampled inputs, so its updates may still encode the offset. We will test whether an attacker can recover $\phi_k$ or user identity from the shared updates, using the re-identification attack protocol of Malcolm et al.
Baselines
Local-only decoder. This is the naive reference. Each device trains alone on its own data with the same architecture and the same number of local steps. Malcolm et al. found it strongest online, but it leaks identity at 100%. Beating it means higher accuracy than it achieves with the same amount of calibration data from the new user.
Plain FedAvg. This is the federated starting point, run on identical device splits, rounds and tuning budget. mixEEG reports FedAvg at 77.0% versus 86.3% for mixEEG on SEED in the domain-generalization setting. That shows how much tailored heterogeneity handling can add, although it is an EEG figure and not a placement-shift figure. We found no reported result on Lacuna for any federated method on CapgMyo or FlexWear-HD under placement shift, so we will establish these numbers ourselves.
FedAvg with placement augmentation. This is the strongest competitor and our most serious threat. Every device augments locally with All Valid Subsets–style shifted sub-grids along the forearm and with circumferential shifts. The aim is a placement-invariant shared model. It is federated only in our experiments, never in the literature. Centrally, All Valid Subsets beat a central-subset model by +6.9% and an all-channel 128-electrode model by +20.7% in intersession CapgMyo tests (Pereira et al.). It gets the same backbone and an equal hyperparameter search budget, with its augmentation range matched to the shift distribution we inject.
Personalized FL with a private input layer. This baseline keeps a full first layer local on each device and federates the rest. Our proposal is the extreme low-parameter case of this design, so this baseline isolates whether the gain comes from structure rather than from privacy of the input layer.
Per-FedAvg and mixEEG-style channel mixup. Both are run as published federated methods, ported to our backbone.
Centralized reference. The EMGBench 2D-CNN pipeline (EMGBench) serves as a non-private upper reference. It reports 83.9% leave-one-subject-out accuracy on FlexWear-HD.
Ablations. Three ablations isolate the key ingredient:
- oracle offsets set to the true injected shift, which gives the ceiling for registration;
- learned offsets that are shared with the server, which tests the privacy cost;
- registration without federation, which tests the value of collaboration.
What counts as success. We will count it a success if the method beats FedAvg with placement augmentation significantly under circumferential shifts, is no worse under small longitudinal shifts, and closes at least half the gap between plain FedAvg and the centralized reference.
Why not just…?
Why not just augment for invariance and run FedAvg? If placement shifts were only small translations within one muscle, invariance would cost little, and this baseline might match us. Rotation is different. Moin et al. found that a 64-electrode array rotated around the forearm failed without explicit feature remapping. Pereira et al. also leave rotation unaddressed. Our argument is that full invariance to circumferential shift erases the information about which muscle compartment is active. We test this directly by sweeping the size of the induced rotation. If the augmentation baseline matches registration at large rotations, our core claim is wrong and we will report that.
Why not just keep the whole input layer private? A private 64×d input layer can represent any remapping. However, it must be learned from each user's few seconds of data, and it is itself a biometric fingerprint of the kind Malcolm et al. exploited. We expect it to win when local data are plentiful and to lose in the few-shot regime. Learning curves over 1–20% of new-user data will show where the curves cross. A real risk is that it wins everywhere. If so, the useful contribution becomes explaining why low-dimensional structure does not help.
Why not just fine-tune a FedAvg model on the new user? EMGBench reports a 44% average gain from 5% of a user's data. This baseline is cheap and strong. We include it applied to both FedAvg and our model, because registration and fine-tuning can be combined. The question is whether starting from a registered decoder reduces the calibration data needed, which we measure as accuracy at fixed seconds of calibration.
Proposed Experiments
Shift-injected EMG benchmarks. We treat each subject in CapgMyo and FlexWear-HD as a device. Every device receives a random longitudinal shift and circumferential rotation, applied by cropping and resampling its physical grid, with held-out subjects receiving unseen offsets. This gives ground-truth offsets for registration error. We report leave-one-subject-out accuracy against shift size, separately for each shift axis.
Real re-application. Moin et al.'s rotated session and the CapgMyo intersession splits test whether the method works on real re-application rather than only on simulated shifts.
Continuous decoding. Malcolm et al.'s offline 14-subject 64-channel cursor-tracking data, with injected shifts, tests velocity decoding error. This matters because their offline federated gains, up to 22%, did not transfer online. Any closed-loop claim therefore needs a later live study, which we treat as out of scope for the first paper.
Privacy. On every federated run we measure re-identification accuracy and offset-recovery error from the shared updates.
Decisive checks. The first decisive check is whether learned offsets track the injected ones and whether accuracy approaches the oracle-offset ceiling. The second is the rotation sweep against FedAvg with augmentation. The third is whether offsets can be recovered from the shared updates. If the third succeeds, keeping offsets local buys no privacy, and we would need noise or secure aggregation on top.
References
- Tackling Electrode Shift in Gesture Recognition with HD-EMG Electrode Subsets
- Nef-Net v2: Adapting Electrocardio Panorama in the wild
- Population Transformer: Learning Population-level Representations of Neural Activity
- Federated Learning in Open- and Closed-Loop EMG Decoding: A Privacy and Performance Perspective
- mixEEG: Enhancing EEG Federated Learning for Cross-subject EEG Classification with Tailored mixup
- EMGBench: Benchmarking Out-of-Distribution Generalization and Adaptation for Electromyography
- An EMG Gesture Recognition System with Flexible High-Density Sensors and Brain-Inspired High-Dimensional Classifier