Spending Relay Power Where the Data Is Good: Secure Over-the-Air Federated Learning for Ambulatory Wearables

Wearables that estimate heart rate from optical pulse signals could learn much better models if many users trained together. Federated learning makes that possible without pooling anyone's raw data: each device trains on its own recordings and sends only model updates. Sending those updates over the air from small, battery-limited devices raises two problems. The first is reliability. In over-the-air computation, all devices transmit at once and the radio channel adds their signals together into the averaged update. This saves bandwidth, but the whole sum is limited by the device with the worst channel. Helper radios called relays can strengthen weak links. The second is secrecy. Anyone listening nearby can intercept the updates, and analog over-the-air sums cannot be encrypted with ordinary ciphers. The same relays could instead broadcast deliberate noise ("cooperative jamming") aimed at an eavesdropper. A relay has one power budget, so every watt spent boosting a weak device is a watt not spent jamming.

This note argues that a third problem, motion artifacts, changes how that power should be split. Arm swing and walking corrupt the optical pulse signal at the same frequencies as the heartbeat. A device whose user is currently walking is therefore computing an update from bad data, and that is also the device whose wireless channel is fluctuating most. If the aggregation weight each device receives reflects the measured quality of its local data, the system no longer has to align every signal to the worst channel. Power that would have rescued a straggler with corrupted data can go to jamming. The core claim is that designing relay forwarding, jamming, and placement against the total error of the aggregated update, which includes per-device data noise and not only channel noise, gives a better accuracy–secrecy trade-off than designing the radio for channel noise and treating data quality as a separate filtering step.

Where this sits in prior work

Two papers already combine relays and federated learning, and both treat the radio as a reliability problem only. Relay-Assisted Cooperative Federated Learning places amplify-and-forward relays to minimize the mean squared error of the over-the-air sum, using alternating optimization. It models no eavesdropper and assumes every device's update is equally trustworthy. Joint Service Pricing and Cooperative Relay Communication for Federated Learning treats multi-hop relaying as a pricing game. We reuse the first paper's two-phase relay protocol and its optimization approach as the method substrate. We change its objective and add a secrecy constraint.

The secrecy literature constrains the design and supplies its tools. Improving wireless physical layer security via cooperating relays shows that a multi-antenna relay can steer jamming into the null space of the legitimate receiver's channel. In its line-of-sight simulation, this keeps positive secrecy at every eavesdropper distance from 10 to 90 m, whereas direct transmission is secure only when the eavesdropper is farther away than the 50 m destination. The same paper shows that the secrecy rate is a U-shaped function of eavesdropper distance, and that the middle of that curve is where jamming is most power-hungry. We reuse its zero-forcing jamming. Relay Placement for Physical Layer Security: A Secure Connection Perspective finds that the best single-relay position is near the source–destination midpoint. We treat that as a hypothesis to test once devices move and their data quality varies. Secrecy and Throughput Performance of Cooperative Cognitive DF Relaying Vehicular Networks warns that relays also create new paths an eavesdropper can intercept. Our secrecy constraint therefore covers the relay's forwarded signal as well as the devices' own transmissions.

On the sensing side, Tiny-PPG labels each sample of a pulse signal as artifact or clean. It reaches 87.4% DICE on PPG-DaLiA with about 20k parameters running on a microcontroller. That makes a per-window data-quality score cheap enough to compute on the device. Q-PPG shows that heart-rate networks can be compressed to a few kilobytes while keeping 4.36 BPM mean absolute error on PPG-DaLiA. A model that small is what makes analog over-the-air aggregation of full updates realistic.

Two papers bear on how quality should be used. FedHealth does federated transfer learning on wearables, but it uses homomorphic encryption and clean motion-sensor activity data. A robust PPG foundation model using multimodal physiological supervision reports that training on noisy real-world segments, anchored to clean ECG, improves robustness. That partly argues against throwing artifact-laden data away. We take it seriously: quality sets a continuous weight, not a binary keep-or-drop decision.

Core Claim and Formulation

Let device $k$ send update $g_k$. The server receives $\hat g = \sum_k a_k g_k + n$. Here $a_k$ is the effective weight that results from device power scaling and relay amplification, and $n$ is the receiver noise left after relaying, whose variance $\sigma_n^2(P_F)$ falls as the relay spends more forwarding power $P_F$. Each local update deviates from the update that clean data would produce, $\bar g_k$, by data noise of variance $\sigma_k^2$. That variance grows with the device's current artifact fraction. Our per-round design problem is

$$\min_{a,\,P_F,\,P_J,\,w}\; \mathbb{E}\Big\|\hat g - \textstyle\sum_k \pi_k \bar g_k\Big\|^2 \approx \sum_k a_k^2\sigma_k^2 + \Big\|\textstyle\sum_k (a_k-\pi_k)\bar g_k\Big\|^2 + \sigma_n^2(P_F)$$

The terms are as follows. $\pi_k$ is the nominal data-size weight that FedAvg would use. $P_J$ is relay jamming power, with $P_F + P_J \le P_R$, the relay's power budget. $w$ is the jamming beam, constrained to the null space of the server channel. The objective is minimized subject to a secrecy constraint: the eavesdropper's signal-to-interference-plus-noise ratio on every device's signal, and on the relayed signal, must stay below $\gamma_E$ in at least a $1-\epsilon$ fraction of fading states. Because the eavesdropper's channel is unknown, the constraint is stated statistically over its possible positions.

The middle term is the bias from deviating from FedAvg weights. Channel-only designs set $\sigma_k^2 = 0$, so their only lever is to match every $a_k$ to $\pi_k$, which forces alignment to the weakest device. When a weak-channel device also has large $\sigma_k^2$, the optimum shrinks its $a_k$, and the forwarding power that device would have needed moves to $P_J$.

Placement is an outer, offline problem. We choose relay positions to minimize the expected value of this per-round optimum. The expectation runs over a mobility model in which each device's position, speed, and artifact level follow its user's activity.

Research Program

We will build three components. The first is a quality estimator on each device. It reports the artifact fraction of each local training window, using either a Tiny-PPG-style segmenter or an accelerometer-energy proxy, and converts that fraction into an estimate of $\sigma_k^2$. The conversion is calibrated by measuring how gradient variance grows with artifact fraction on held-out subjects. Only this one scalar per round leaves the device in the clear.

The second is the per-round solver. It extends the alternating optimization of the relay-assisted federated learning paper with the zero-forcing jamming of Dong et al. and the secrecy chance constraint.

The third is a coupled mobility-and-channel simulator. PPG-DaLiA activity labels drive each subject's trajectory: sitting is static, walking moves at a typical gait speed. Positions feed a path-loss model with exponent 3.5, matching the Dong et al. simulation, plus small-scale fading whose Doppler rises with speed. The coupling between motion and channel is synthetic, and we say so. The final experiment varies its strength, so the claims do not hinge on it.

The analytical question is when quality-aware weighting moves the optimal relay away from the geometric midpoint. We predict that the optimum shifts toward where users are stationary and producing usable data. If the midpoint stays optimal under realistic mobility, that is a reportable negative result.

Baselines

The naive reference is direct over-the-air FedAvg with no relay and no jamming. Under the secrecy constraint, its admissible operating region should collapse whenever the eavesdropper is closer than the server, as Dong et al. predict. It also gives the untouched accuracy we must improve on.

The strongest existing competitor is a decoupled pipeline. It runs the mean-squared-error-optimal relay design of Relay-Assisted Cooperative Federated Learning with FedAvg weights. Power left after meeting its alignment target goes to zero-forcing jamming from Improving wireless physical layer security via cooperating relays. Relay positions follow the midpoint rule of Relay Placement for Physical Layer Security. It is the strongest competitor because it is exactly what one would build by stacking existing papers, so any gain we show must come from the coupling itself. No reported result found on Lacuna for relay-assisted or secure federated heart-rate estimation on PPG-DaLiA. The file also gives no numeric aggregation error from the relay-assisted federated learning paper. We therefore run this baseline ourselves, with the same relay budget, antenna count, secrecy target, and number of rounds, and with its optimizer given the same iteration budget as ours.

We count it as beaten under two conditions at equal secrecy outage. Either our method reaches lower final leave-subject-out heart-rate error, or it sustains a higher secrecy target (for instance the 1 b/s/Hz level used by Dong et al.) at equal error. The gap must hold across at least three eavesdropper regions on the U-curve.

For accuracy ceilings we report centralized training of the same Q-PPG-sized network on PPG-DaLiA. Q-PPG reports 4.36 BPM mean absolute error for its centrally trained model. The robust foundation model reports 7.78 BPM under its own protocol. We also report error-free digital federated training with quality weighting. That ceiling ignores bandwidth and secrecy, and it bounds how much the radio is costing us.

The ablations isolate the key ingredient. One keeps quality weighting but freezes the relay design at the decoupled solution. A second keeps the joint radio design but sets every $\sigma_k^2$ to zero. A third replaces the segmenter-based quality score with accelerometer energy.

Why not just…?

Why not just drop artifact-contaminated windows on the device and then run the decoupled relay pipeline? This is the real threat. Hard dropping is a binary version of our weights, and if most windows are either clearly clean or clearly corrupt, it could recover nearly all of the gain. Two considerations argue against assuming it will. The foundation-model result suggests noisy segments still carry useful structure. And dropping removes the walking subjects precisely when their data differs most, which harms generalization to motion. Neither consideration settles the question, so we run the drop-then-decouple pipeline as a named arm. If it matches our method, we report that the coupling reduces to "schedule only clean devices" and that continuous weights add nothing.

Why not just encrypt the updates? Over-the-air aggregation relies on analog superposition of the update values. Ciphertexts are pseudo-random bit strings that do not add meaningfully over the air, so encryption forces orthogonal digital uploads. The homomorphic route used by FedHealth also inflates payloads well beyond what microcontroller budgets like Q-PPG's 1.79 mJ per inference suggest is affordable. The digital-ceiling arm measures the accuracy that encryption-plus-orthogonal-uploads would buy. Reporting its airtime makes the cost of that accuracy explicit.

Why not just use robust aggregation, such as a coordinate-wise median, to ignore bad updates? A median requires individual updates to be decoded separately, which again rules out over-the-air summation. It also discards the side information the device already has about why its update is bad. Within the analog setting, the closest analogue is the weights-only ablation. If that ablation captures most of the gain, the radio-side coupling is unnecessary, and we would say so.

Proposed Experiments

We treat PPG-DaLiA's 15 subjects as 15 devices and evaluate leave-one-subject-out heart-rate mean absolute error. The model is a Q-PPG-sized temporal convolutional network, small enough for full-vector analog upload. Wireless conditions come from the simulator above, with one or two multi-antenna relays (three to six antennas, following Dong et al.). The eavesdropper is placed in the near, middle, and far regions of the U-curve, and its channel is drawn statistically rather than assumed known. We report heart-rate error against rounds, secrecy outage probability, and how relay power is split between forwarding and jamming. Results are broken down by activity class, so we can see whether the walking segments, where camera-based heart-rate estimation errors on MMPD exceed 25 BPM (MMPD), are where the gains or losses concentrate.

Three checks are decisive. First, at matched secrecy outage, the joint design must beat both the decoupled pipeline and drop-then-decouple on held-out error. Second, the power-split logs must show the claimed mechanism: jamming power should rise in rounds where the weakest-channel device is also the most artifact-heavy. Third, we vary the strength of the coupling between motion and channel quality from zero to strong. If the gain survives with zero coupling, it comes from quality weighting alone. If the gain appears only under strong coupling, the claim narrows to environments where walking users really do have the worst links. That would be a precise and useful conclusion, but a smaller one than the title promises.