Consider federated learning (many devices jointly train a model while their raw data stays local) on sensors powered by sun, wind, or heat. These devices can only take part when their small energy store allows it. That creates two problems. The first concerns accuracy. Energy-rich devices take part more often, so the model drifts toward their data. The standard fix is to scale each device's update by the inverse of how often it can take part. That fix breaks differential privacy (a guarantee that adds calibrated noise so no single record can be inferred), because a device that harvests once every twenty rounds sends an update twenty times larger and needs twenty times more noise. The second problem is confidentiality. The schedule of who takes part, and when, follows each device's battery. A curious server can read that schedule to learn where the device sits, what its weather is like, and how its owner behaves. Encrypting the updates does not help, because the leak is in the timing.
We propose one mechanism that addresses both problems. Every device takes part in each round with the same fixed probability $q$, decided by a private coin flip. The battery and supercapacitor absorb the mismatch between the coin's demand and the irregular harvest. Because every device has the same rate, no inverse scaling is needed and the privacy sensitivity stays at the clipping bound. Because the coin is independent of energy, the participation schedule carries no information about the device's physical state. The one exception is an outage, when the coin says "go" and the battery is empty. Hybrid harvesting is what makes this practical. Complementary sources such as solar with wind, or solar with thermal, lower the long-run variance of the harvest. We argue below that the outage probability falls exponentially as variance falls. A modest buffer can therefore support a usefully high $q$ with a provably small leak.
Core Claim
For federated learning on energy-harvesting clients under a semi-honest server, we claim three things. A semi-honest server follows the protocol exactly but logs and analyses everything it sees. The first claim is that energy-decoupled fixed-rate participation keeps the unbiasedness of inverse-probability weighting at the sensitivity of plain clipping. The second is that it bounds the information the participation schedule leaks about each client's energy process by that client's outage probability over the training horizon. The third is that this outage probability, and therefore the reachable privacy-utility frontier, is set by the buffer size and the long-run variance of the harvest. That variance is exactly what multi-source harvesting reduces. The scheme deliberately wastes energy at energy-rich clients, which run below their natural rate. This mirrors the finding in Increasing Smart Meter Privacy Through Energy Harvesting and Storage Devices that perfect privacy of physical load requires wasted energy. That paper treated a battery as a filter on household electricity demand. We move the same idea from the electrical layer to the timing of machine-learning messages.
Mechanism and Why Hybrid Sources Matter
Let $H_i(t)$ be the energy client $i$ harvests in round $t$, $e$ the energy cost of one participation, $B_i(t)$ the stored energy, and $K$ the buffer capacity. Each round, the client draws a private coin $Z_i(t)\sim\mathrm{Bernoulli}(q)$ and participates only if $Z_i(t)=1$ and $B_i(t)\ge e$. An outage is a round where $Z_i(t)=1$ but $B_i(t)<e$.
Outside outages, the participation sequence is independent and identically distributed Bernoulli($q$) whatever the weather. So for any test a semi-honest observer runs on client $i$'s participation transcript $P_i$ over $T$ rounds, the gap between two energy traces $h$ and $h'$ is bounded by the outage probabilities:
$$\mathrm{TV}\big(P_i \mid h,\; P_i \mid h'\big) \le \Pr[\text{outage in } T \mid h] + \Pr[\text{outage in } T \mid h'].$$
Here $\mathrm{TV}$ is total-variation distance, which upper-bounds any adversary's advantage in telling the two apart. The buffer acts like a queue with drift $\mu=\mathbb{E}[H]-qe$ and long-run variance $s^2$. The long-run variance includes the autocorrelation of the harvest, such as the daily solar cycle. The standard diffusion approximation for the probability that such a queue drains gives:
$$\Pr[\text{outage per round}] \approx \exp\!\big(-2K\mu/s^2\big).$$
This expression shows the trade-off directly. At a fixed mean harvest, complementary sources shrink $s^2$, so they allow a smaller buffer or a higher $q$ at the same leakage. The approximation needs empirical validation on real, non-stationary traces. Checking it is part of the program.
For the training data itself, the uniform $q$ means the server averages updates that are clipped at norm $C$ and given equal weight. Secure aggregation lets the server see only the sum. Distributed Gaussian noise is added under the semi-honest model. Standard per-record accounting then applies, including amplification from each client's local minibatch sampling, which the server cannot see. We deliberately do not count environmental randomness as privacy noise. The server can know the weather, and amplification arguments such as those of Dong et al. require randomness that the adversary does not see and that does not depend on the data. Physical entropy in harvesting fails both conditions. This rules out the tempting "free noise from the sun" design.
Research Program
The first strand is analytical. We will formalise the transcript leakage bound and the outage approximation for autocorrelated, multi-source harvest processes. We will then derive how utility scales with $q$ under distributed DP. Expected participants per round are $qn$, so the noise on the average scales as $\sigma C/(qn)$, where $n$ is the number of clients and $\sigma$ is the noise multiplier. The result should be a single design curve linking $q$, $K$, the source mix, the leakage bound, and accuracy.
The second strand is measurement. On the KRATOS platform we will record harvest traces from solar alone, thermal alone, and both together through the BQ25570 power management chip into its supercapacitor. KRATOS uses an MSP430 microcontroller with ferroelectric RAM, draws 1.83 µW asleep, and draws 240 mW when transmitting over LoRa. We will run the coin-and-buffer policy on the device to measure real outage rates. The file contains no public solar–wind traces, so the solar–wind case will use synthetic complementary processes, and we will state this limitation openly.
The third strand is learning experiments. Training runs in simulation, driven by those traces, because no deep model trains on an MSP430. The energy cost per round $e$ is calibrated from the measured radio power. The fourth strand is attacks. We will build a semi-honest server attacker that tries to infer each client's energy group, source type, and weather from participation timing. We will compare its success against the bound.
Baselines
The naive reference is energy-agnostic DP-FedAvg. Clients participate whenever they have energy, updates are clipped at a common $C$, and Gaussian noise is added, with no reweighting. In the non-private setting, Energy-Harvesting Distributed Machine Learning reports that its energy-agnostic benchmark reaches "only 64% accuracy" on CIFAR-10, against roughly 80% for its method. We expect the private version to be biased in the same way and to leak heavily through timing.
The strongest competitor on utility is that paper's unbiased inverse-probability scaling (its Algorithm 1). We will add DP in its two honest forms. In the first, clients clip after scaling, which keeps sensitivity at $C$ but introduces clipping bias. In the second, clients clip before scaling, so sensitivity is $C/\beta_{\min}$ and noise grows to match. Without privacy this method reaches "approximately 80%" on the 40-client, four-group CIFAR-10 setup, matching distributed SGD with unlimited power. No reported result found on Lacuna for any energy-harvesting federated method under formal DP, so we will produce that number ourselves. We will also include battery-aware Myopic scheduling with the adaptive learning rate $\eta_t\propto\sqrt{n_t}$ from On Federated Learning with Energy Harvesting Clients, with DP added. Because it schedules by battery level, it is the most leakage-prone strong baseline.
For timing leakage, the reference is the heuristic "mingling ratio" of Clustered Federated Learning for Energy-Harvesting Smart Meters in P2P Energy Trading. It injects decoy clients into clusters and reports that habit-profiling accuracy fell "from ~89% to ~34%" at a mingling ratio of 0.4, without a formal guarantee.
Every method gets the same data partitions, energy traces, round budget, accounting code, and target $(\epsilon,\delta)$. Each gets an equal hyperparameter sweep over $C$, $\sigma$, and learning rate. We will also run every baseline with the adaptive clipping of Adaptive Clipping for Differential Private Federated Learning in Interpolation Regimes, so that a gain from our method cannot be explained by better clipping.
Two ablations isolate the key ingredients. The first keeps the fixed-rate coin but uses per-group rates $q_g$ with $1/q_g$ weighting. This tests whether uniformity is what preserves sensitivity, and whether publishing $q_g$ leaks group identity. The second compares single-source against hybrid traces at the same mean harvest. This tests whether variance reduction, rather than extra total energy, drives the gains. Our method beats the baselines if, at equal $(\epsilon,\delta)$, it matches or exceeds the best private baseline's accuracy while cutting the timing attacker's advantage to within the measured outage bound.
Why not just…?
Why not just use inverse scaling and clip after scaling? This keeps sensitivity at $C$, but it clips hardest exactly the energy-poor clients that scaling was meant to amplify, so the bias returns through the back door. Adaptive Clipping for Differential Private Federated Learning in Interpolation Regimes shows that a poorly chosen clipping radius distorts the gradient direction. The approach also does nothing about timing leakage. It is a genuine utility competitor, and the head-to-head under adaptive clipping will settle it.
Why not just set $q$ to the poorest client's harvest rate and use ordinary sampling, with no buffer analysis? This is a real threat, because it is essentially our mechanism without the theory. The difference is that without sizing the buffer against autocorrelated harvests, such as night-long solar gaps, outages will be frequent. Outage timing then leaks exactly the energy state the scheme is meant to hide. If the measured outage rates turn out negligible even with tiny buffers, our contribution reduces to the leakage bound and the attack evaluation, and we will report that plainly.
Why not wait until the battery is full, then run a standard semi-honest protocol? Energy-Harvesting Distributed Machine Learning reports that its wait-for-all benchmark reached "only 52% accuracy within the 1,000-iteration window." Federated Learning With Energy Harvesting Devices: An MDP Framework identifies a persistent error floor caused by intermittent participation. Threshold wake-ups also make participation a direct readout of when the battery filled. We include this policy as a baseline, both to confirm the accuracy loss and to measure its leakage.
Why not treat the physical randomness of harvesting as free DP noise? Amplification needs randomness the adversary cannot see and that is independent of the data, as in Dong et al.. Weather is public, and the server can model it, so the physical randomness of harvesting meets neither condition. Our design spends a private coin instead, and uses the physical randomness only through its variance, which sets the buffer size.
Proposed Experiments and Decisive Checks
The main learning benchmark replicates the 40-client, four-group CIFAR-10 setup of Güler and Yener, with arrival rates ranging from every step to once every twenty steps. It adds non-i.i.d. label splits and trace-driven arrivals, and FashionMNIST serves as a cheaper second task. We report accuracy at $\epsilon\in\{1,4,8\}$. As a centralized, non-federated reference point, PLRV-O reports that the standard Gaussian mechanism reaches 83.93% on CIFAR-10 with a vision transformer at $\epsilon\approx0.5$. This marks an upper range that federated energy-limited training should not be expected to reach.
The first decisive check is utility. At equal $\epsilon$, fixed-rate participation should beat clip-before-scale inverse weighting clearly for the energy-poorest group, and match or beat clip-after-scale overall. If clip-after-scale with adaptive clipping wins, the utility part of the claim fails.
The second decisive check is leakage. Across buffer sizes, a classifier that infers energy group and source type from participation timing should have measured advantage no larger than the outage bound. Myopic and threshold policies should leak near-perfectly.
The third decisive check is the hybrid effect. On KRATOS solar, thermal, and combined traces, the log-outage rate should fall roughly linearly in $K\mu/s^2$, as the diffusion approximation predicts. At matched mean harvest, the hybrid traces should support a higher $q$ at the same leakage, and therefore higher accuracy at the same $\epsilon$. If complementarity buys nothing beyond total energy, the hybrid-harvesting part of the claim fails, and we would report the mechanism as source-agnostic.
References
- Increasing Smart Meter Privacy Through Energy Harvesting and Storage Devices
- Leveraging Randomness in Model and Data Partitioning for Privacy Amplification
- KRATOS: An Open Source Hardware-Software Platform for Rapid Research in LPWANs
- Energy-Harvesting Distributed Machine Learning
- On Federated Learning with Energy Harvesting Clients
- Clustered Federated Learning for Energy-Harvesting Smart Meters in P2P Energy Trading
- Adaptive Clipping for Differential Private Federated Learning in Interpolation Regimes
- Federated Learning With Energy Harvesting Devices: An MDP Framework
- PLRV-O: Advancing Differentially Private Deep Learning via Privacy Loss Random Variable Optimization