Battery-free medical sensors, such as cardiac patches and small implants, could in principle train a shared diagnostic model together without any patient's raw signals leaving the body. In this setting, called federated learning, devices train locally and send only model updates to an aggregator. The difficulty is energy. Existing work on federated learning with energy-harvesting devices assumes the energy comes from ambient sources like sunlight, so a device takes part in training whenever its battery happens to be full. That makes participation bursty and correlated over time. It also quietly breaks one of the most useful tools in private training: privacy amplification by subsampling. This is the fact that if each patient's device joins a training round only by a secret coin flip, the formal privacy loss per round shrinks roughly in proportion to the sampling rate. A device that participates "every time its battery refills" has not been sampled at random, and the standard accounting no longer applies.
Simultaneous wireless information and power transfer (SWIPT) means that a base station sends data and electrical energy on the same radio wave, and the receiver splits the incoming power between a decoder and a rectifier. SWIPT changes the situation because the aggregator now decides who receives energy. Our central idea follows directly. The aggregator should first draw a secret, data-independent random sample of patients, and only then beam energy to the sampled devices, together with the global model, so they can complete the round. Participation becomes memoryless again, and amplification can be recovered. Unsampled patients' bodies receive no unnecessary radio-frequency exposure. We propose to formalize this "sample-then-power" protocol under tissue-safety limits, derive per-patient privacy accounting when delivery can fail for far-away devices, and test whether it gives better ECG diagnostic accuracy at a fixed patient-level privacy budget than the energy-harvesting schedulers and simpler alternatives discussed below.
Core Claim
In wirelessly powered federated learning, the order "random sample, then deliver energy" yields participation that is independent across rounds and independent of patient data. The order used implicitly by ambient-harvesting methods, "accumulate energy, then participate when able," does not. The first order supports standard subsampled differential privacy accounting, with a per-patient sampling probability set by the power budget and the radio channel. The second supports at best accounting without amplification. We claim this difference is large enough to matter clinically: at the same patient-level privacy budget $(\epsilon, \delta)$, sample-then-power should give a measurably higher diagnostic macro AUC on 12-lead ECG classification than battery-driven scheduling. It should also use less total radio-frequency energy and tissue exposure than charging every device and sampling afterwards.
The idea we deliberately do not pursue is the most obvious one at this intersection. Gap analyses of this literature suggest treating the power-splitting ratio $\rho$ as a privacy knob, because diverting power to the harvester raises the decoder's bit error rate, and Free Privacy Protection for Wireless Federated Learning shows that channel bit errors can serve as differential-privacy noise. The flaw is that $\rho$ acts at the device's receiver on the downlink, which carries the global model. Corrupting the downlink protects no patient's data; it only degrades the model the device trains on. Uplink channel noise as privacy noise is a separate mechanism, and it requires trusting that the aggregator cannot improve its own receiver. In our design, $\rho$ remains purely an energy-versus-model-fidelity control. Privacy comes from sampling plus calibrated noise.
Why the Gap Is Real
Three papers form the energy-harvesting federated learning line, and all three motivate this proposal by sharing one assumption. Federated Learning over Energy Harvesting Wireless Networks schedules users by channel gain and battery queue under massive multiple-antenna beamforming, and reports about 93% accuracy on an MNIST subset. On Federated Learning with Energy Harvesting Clients shows that a "myopic" policy, which schedules the fullest batteries first, beats greedy and round-robin scheduling. It proves an $O(1/\sqrt{n_{min}T})$ convergence rate despite participation that is correlated over time. Over-the-Air Federated Learning with Energy Harvesting Devices adds a cooldown multiplier so energy-rich devices do not dominate the model. All three treat energy arrival as an exogenous random process, and none carries a privacy guarantee. Their schedulers are exactly the battery-driven participation rules whose privacy accounting we argue is weak. We reuse their convergence machinery but replace the participation process.
On the SWIPT side, power splitting and its trade-off between data rate and harvested energy (Wireless Information and Energy Transfer in Multi-Antenna Interference Channel; Data and energy integrated communication networks) supply the delivery model. Dynamic power splitting under varying battery and channel states (Performance Analysis of SWIPT with Ambient RF Energy Harvesting) supplies the per-round control. Achieving Covert Communication in Large-Scale SWIPT-Enabled D2D Networks is the closest SWIPT-plus-security work. It hides the existence of transmissions from wardens in a static snapshot, with no learning and no differential privacy, so it constrains our channel model but not our claim. Safety enters through specific absorption rate limits on tissue heating (Specific Absorption Rate-Aware Multiuser MIMO Assisted by Fluid Antenna System). These limits cap the energy any patient can receive per round. That cap is what makes delivery fail for distant devices.
On the clinical side, A Federated Learning Framework for Healthcare IoT devices splits an ECG network between device and cloud and sparsifies traffic. It reaches within 2% of centralized accuracy on PhysioNet 2017 while cutting traffic from 5.45 GB to 12.8 MB per round across 64 devices, but it has no formal privacy noise and assumes ordinary batteries. Membership inference against clinical models (Exploring Membership Inference Vulnerabilities in Clinical Large Language Models) motivates insisting on formal guarantees rather than relying on architectural splitting alone.
Research Program
The system is a single aggregator with a multi-antenna transmitter serving $N$ patient-worn or implanted ECG devices. Each device stores almost no energy, so it can complete local training only with energy delivered in that round. At round $t$, the aggregator draws an independent Bernoulli sample in which patient $i$ is included with target probability $q_i$. It keeps the sample secret from everyone except the sampled devices. It then solves a per-round beamforming and power-splitting problem: deliver at least the computation-plus-uplink energy $E_{req}$ to each sampled device, subject to per-patient absorption caps and total transmit power, while keeping the downlink model decodable. A device whose channel cannot support delivery under the cap simply does not finish. Because failure depends only on the channel and the sample, never on the ECG data, the effective participation probability is
$$ \tilde q_i = q_i \cdot \Pr\left[E_i^{deliv}(\mathbf{h}_i, \rho_i) \ge E_{req} \;\middle|\; \text{SAR}_i \le \text{SAR}_{max}\right], $$
where $\mathbf{h}_i$ is patient $i$'s channel and $\rho_i$ the fraction of received power sent to the harvester. Participation stays independent across rounds and data-independent. Standard subsampled-Gaussian Rényi accounting therefore applies per patient with rate $\tilde q_i$. Since a lower rate only strengthens privacy, a conservative bound can use the target $q_i$.
Three analyses follow. The first derives the per-patient privacy loss after $T$ rounds as a function of $\tilde q_i$, the update clipping norm, and the noise scale, and compares it with the best available accounting for myopic battery-driven participation. That comparison counts only actual participations and gives no amplification. The second adds a fairness layer. Distant patients have lower delivery probability, so they participate less, which leaves them with less model influence but tighter privacy. We optimize the $q_i$, raising the sampling rate for far patients, and the beam allocation so that the participation-weighted contribution is balanced across patients. This is the analogue of the cooldown correction of Aygün et al., achieved through sampling rather than reweighting. The third adapts the convergence bound of Shen et al. to independent heterogeneous participation, which is a simpler case than their correlated one, with an extra error term for the downlink model corruption controlled by $\rho_i$.
Local computation is bounded by intermittent-power realities. Revisiting DNN Training for Intermittently Powered Energy Harvesting Micro Computers reports roughly a 14% training-time overhead for checkpointing under power failure. We therefore use a device-side front end in the style of Yuan et al.: the device runs only a shallow feature extractor, and the deeper layers live at the aggregator. Under sample-then-power, the device needs no checkpoints, because energy for the round arrives before computation starts.
Baselines
The naive reference is DP federated averaging driven by the ambient-harvesting participation process. Devices join whenever their simulated battery fills. Because participation is not random sampling, privacy is accounted by plain composition over each patient's actual participations. This shows what a practitioner gets by bolting standard differential privacy onto existing energy-harvesting federated learning.
The strongest existing scheduling competitor is the myopic fullest-battery policy of On Federated Learning with Energy Harvesting Clients, with its square-root-scaled learning rate. It is strongest because it was shown to beat greedy and round-robin scheduling and comes with a convergence guarantee. We also run the channel-and-battery-aware scheduler of Federated Learning over Energy Harvesting Wireless Networks, the strongest method that exploits channel knowledge. Both receive the same SWIPT energy model, the same total transmit-power and absorption budget, the same model, the same per-patient $(\epsilon, \delta)$, and the same hyperparameter search budget over clip norm, learning rate, and noise multiplier. Each method's noise is calibrated by its own valid accounting.
The upper reference is a battery oracle: ordinary devices with ample energy running Poisson-sampled DP federated averaging. Sample-then-power should approach this oracle, and the gap measures the cost of wireless powering and failed deliveries.
On accuracy, the reference point is centralized, non-private, non-federated training. Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL reports a macro AUC of approximately 0.939 on diagnostic statements for CNN ensembles, with wavelet-plus-shallow-network baselines around 0.87–0.90. For single-lead PhysioNet 2017, Yuan et al. report under 2% accuracy loss relative to centralized training without formal privacy. We found no reported result on Lacuna for differentially private federated training on PTB-XL or PhysioNet 2017. We will therefore report our own privacy-free federated number as an intermediate reference.
The key ablation is "power-then-pick," which isolates the ordering. The aggregator spends the same per-round energy budget on the same beamforming optimizer, but picks among devices that already hold enough energy. This keeps everything identical except whether the random draw precedes energy delivery. Sample-then-power beats the field if, at matched $(\epsilon, \delta)$ and matched total radiated energy, it improves PTB-XL diagnostic macro AUC over the best battery-driven scheduler by more than the seed-to-seed spread, and if the ablation shows that the gain disappears when the ordering is reversed.
Why not just…?
Why not just give patients rechargeable batteries and use ordinary DP federated averaging? For wearable patches that can be recharged nightly, this is the right answer, and our battery oracle encodes it. The proposal targets devices where recharging means surgery or poor patient compliance, a constraint the file grounds only qualitatively (Exploring older peoples' attitudes and preferences around the use of their healthcare information). We will report the gap to the oracle honestly. If the gap is small, the conclusion is that wireless powering costs little privacy when ordered correctly, which is still useful. We do not claim batteries are worse where they are practical.
Why not just charge every device continuously and then sample randomly among the charged ones? This also restores independent sampling, and it is the most serious threat to the idea. Its cost is that every patient receives radio-frequency exposure every round, even though only a fraction $q$ of them train, and far patients may never charge fully under absorption caps. We include it as a baseline at the same per-patient absorption limit. Sample-then-power must match its accuracy at markedly lower total radiated energy and cumulative exposure per patient, and it should cover more distant patients. If charge-all matches on both accuracy and exposure, for example because absorption caps never bind at realistic distances, then our contribution reduces to the accounting analysis, and we will say so.
Why not just treat battery-driven participation as if it were Bernoulli sampling at its average rate? This is unsound on paper, but it may be harmless in practice. We will test it directly by inserting canary patients and running membership-inference audits against models trained under myopic scheduling with this optimistic accounting. If audits detect no excess leakage, the practical case for reordering weakens. If they do detect leakage, the case for reordering strengthens.
Why not just use a private training method that does not need random sampling? Tree-aggregation and correlated-noise training methods give guarantees under arbitrary participation patterns, so in principle they could absorb battery-driven schedules. The file contains no results for them in this setting. We will implement one as an additional baseline under the myopic schedule, at the same budget. If it closes the gap, the honest conclusion is that sampling order matters only for amplification-based accounting, and we will scope the claim to that.
Proposed Experiments
The main benchmark is PTB-XL (PTB-XL, a large publicly available electrocardiography dataset) at 100 Hz, since 500 Hz brought no significant diagnostic gain in the benchmark study. We use the dataset's patient-stratified folds, assign each of the 18,885 patients to a simulated device, and target patient-level differential privacy. PhysioNet 2017 single-lead segments, with the 74,275/13,107 train/test split used by Yuan et al., serve as the wearable setting. The model is a compact one-dimensional residual CNN, split between device and aggregator. Devices are placed at random distances from a multi-antenna aggregator using the same channel model in every run. Energy is converted through a saturating nonlinear rectifier curve and capped by per-patient absorption limits. Required per-round energy comes from operation counts of the device-side layers.
We report diagnostic macro AUC against the per-patient $\epsilon$ at fixed $\delta$, sweeping $\epsilon$ from strict to loose. We also report total radiated energy and cumulative per-patient exposure, the fraction of patients who ever participate, macro AUC on the most distant patient quintile, and membership-inference audit advantage on canaries.
The decisive checks are three. First, at matched $\epsilon$ and matched radiated energy, sample-then-power must beat the myopic and channel-aware schedulers on macro AUC, and the power-then-pick ablation must lose that advantage. Second, sample-then-power must match charge-all-then-sample on accuracy at substantially lower exposure. Third, the canary audits must confirm that our accounting is not loose in the other direction. Failing the first check would falsify the core claim, and failing the second would reduce it to a bookkeeping result.
References
- Free Privacy Protection for Wireless Federated Learning: Enjoy It or Suffer From It?
- Federated Learning Over Energy Harvesting Wireless Networks
- On Federated Learning with Energy Harvesting Clients
- Over-the-Air Federated Learning with Energy Harvesting Devices
- Wireless Information and Energy Transfer in Multi-Antenna Interference Channel
- Data and energy integrated communication networks
- Performance analysis of simultaneous wireless information and power transfer with ambient RF energy harvesting
- Achieving Covert Communication in Large-Scale SWIPT-Enabled D2D Networks
- Specific Absorption Rate-Aware Multiuser MIMO Assisted by Fluid Antenna System
- A Federated Learning Framework for Healthcare IoT devices
- Exploring Membership Inference Vulnerabilities in Clinical Large Language Models
- Revisiting DNN Training for Intermittently-Powered Energy-Harvesting Micro-Computers
- Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL
- Exploring older peoples' attitudes and preferences around the use of their healthcare information
- PTB-XL, a large publicly available electrocardiography dataset