Smart meters report household electricity use every few seconds, and that stream is detailed enough for an eavesdropper to infer which appliances run and when. This inference is called non-intrusive load monitoring (NILM): an algorithm reads only the whole-house total and separates it into individual appliance traces. The main published defenses are rechargeable batteries that reshape the power drawn from the grid, added statistical noise, and rescheduling appliances. Each is designed and scored against a scalar view of the data, meaning one active-power number per time step, modelled as independent draws or a simple Markov chain. Real meters record more than that. Active power, reactive power, and time together trace a curve, and appliances differ in the shape and ordering of that curve: a compressor's reactive power may lead its active power, while a heater has almost none. We propose to test whether defenses certified on scalar active power still leak through this multichannel geometry. We also propose to build a cheap, interpretable leakage probe that can be run across many defenses and houses.
The tool is the path signature. It treats a window of readings as a continuous curve and summarizes it with iterated integrals. The first level gives each channel's net change. The second level gives signed areas between pairs of channels, and those areas record which channel moved first. A known result says any continuous function of the curve is approximately linear in these features. That makes a linear model trained on signatures a principled, low-capacity stand-in for a flexible attacker. Our claim has two parts. First, battery and noise defenses that drive scalar-power leakage toward zero will leave appliance activity detectable by a linear readout of level-2 signatures of (active power, reactive power, time). Second, a battery policy whose objective penalizes signature-level distinguishability closes that gap at equal battery capacity.
Grounding in prior work
Three results motivate the gap. Farokhi's Fisher-information bound proves that an attacker's ability to locate appliance start times depends on how much the appliances' power derivatives overlap, not on their power levels. The bound is limited to scalar profiles and first derivatives. Level-1 signature terms are exactly integrated derivatives, and level-2 terms are their natural multichannel, order-sensitive extension, so we read our proposal as asking whether Farokhi's insight survives once more than one channel is available.
The battery-masking line casts battery control as a Markov decision process. It quantifies privacy as the mutual information between true demand and grid intake. Li et al. do this under Markov or i.i.d. demand, and Erdemir et al. extend it to renewables and finite storage with "store-and-hide" and "best-effort" policies. Both measure leakage only in scalar active power. Wang et al. give differential-privacy bounds for NILM, but they model appliances as binary switching events. The file itself flags this as unable to represent multi-state, continuous trajectories.
The method comes from the signature literature. We rely on the linearization property described by Compagnoni et al. and Bleistein et al., which justifies linear readouts. We rely on the order-encoding role of level-2 terms shown in Kormilitzin et al.. We use the invariance to irregular sampling described by Moreno-Pino, which matters because smart-meter logs have dropouts. Because that invariance discards duration, we append an explicit time channel; appliance duration is itself a privacy-relevant cue.
The deep NILM models constrain the evaluation. Convolutional Seq2Seq NILM and STAFN are the strongest attackers, so any "leakage" claim must be checked against them rather than only against our own probe.
No paper in the file applies signatures to NILM or to smart-meter privacy. The nearest signature work, CPDA, aligns signature distributions for domain adaptation, a different goal. We deliberately do not propose "signatures as a better disaggregator." Deep models already disaggregate well, and a lighter model would not change what is known about privacy risk.
Core Claim
Let $x_{[s,s+w]}$ be a window of length $w$ of the three-channel grid-intake path (active power, reactive power, time). Let $S^{\le m}(\cdot)$ be its log-signature truncated at level $m$. The probe is an $\ell_2$-regularized logistic readout $f(x)=\sigma(\beta^\top S^{\le m}(x))$ that predicts whether a target appliance is active in the window. We define signature leakage as the probe's held-out area under the ROC curve (AUC) minus 0.5.
The empirical claim is that, under scalar-optimal battery policies, level-2 leakage stays well above level-1 and scalar-only leakage for motor-driven and multi-phase appliances. The design claim concerns a signature-aware battery policy. This policy minimizes the energy cost of grid intake plus $\lambda$ times the squared maximum mean discrepancy between the mean signature features of intake windows with and without the target appliance active:
$$\min_{\pi}\; \mathbb{E}_\pi[\text{cost}] + \lambda \big\| \mathbb{E}_\pi[S^{\le 2}(y)\mid a=1] - \mathbb{E}_\pi[S^{\le 2}(y)\mid a=0] \big\|^2 .$$
Here $\pi$ is the battery charge/discharge policy (for inverter-coupled storage, including reactive-power injection), $y$ is the resulting intake window, $a$ is the appliance-activity indicator, and $\lambda$ trades cost against privacy. Driving this gap to zero makes the two classes identical in mean signature features. Because signatures are approximately universal features, that removes what any linear-in-signature attacker can exploit. It does not remove everything a nonlinear deep attacker might find, which is exactly what the experiments test.
Research Program
The first stage builds the audit. We reproduce the Li et al. Markov-optimal policy, Erdemir et al.'s store-and-hide and best-effort policies, Laplace noise at several privacy budgets following Wang et al., and Farokhi's schedule-shifting defense. Each is applied to UK-DALE whole-house data, where active and apparent power allow reactive power to be derived, and to REDD. For every defended trace we fit three probes on identical windows: a scalar-P probe, a level-1 signature probe, and level-2 and level-3 signature probes on (P, Q, t). We also retrain the deep attackers on the defended data.
The second stage attributes leakage. Linear readouts over named iterated integrals let us report which channel pairs and which orderings carry the signal, for example the P–Q area term for fridges versus heaters. This turns a single leakage number into guidance for the defense.
The third stage trains the signature-aware policy. We use approximate dynamic programming, or policy-gradient optimization over a battery with rate limits and round-trip losses, a constraint the file notes prior work ignores. We then re-run the full audit on its output, including deep attackers the policy never saw.
Baselines
The naive reference point is the undefended meter attacked by a scalar-P probe. This anchors what "no privacy" looks like and what cheap attackers already achieve. The literature sets the strongest disaggregation numbers on undefended data. Convolutional Seq2Seq NILM reports cross-house REDD mean absolute error (MAE, train House 2, test House 1) of 32.0 W for the fridge and 12.8 W for the dishwasher. STAFN reports overall MAE of 19.58 and signal aggregation error of 0.0568 on UK-DALE. We will reproduce these on the same splits before trusting our pipeline. No reported result was found on Lacuna for any deep NILM attacker operating on battery- or noise-defended traces, so those numbers are produced here.
The strongest competitor on the privacy question is a deep attacker retrained on each defended dataset. STAFN and Seq2Seq are both multichannel-capable once Q is added as an input. We give them the same windows, the same train/test households, and a matched hyperparameter search budget, with GPU time recorded. The signature probe "beats" them in a specific sense: its ranking of defenses by leakage agrees with the deep attackers' ranking (Kendall's tau of at least 0.8 across roughly 20 defense configurations) at a small fraction of the compute. We do not claim it will extract more information than they do.
For the defense, the baselines are store-and-hide and best-effort at equal usable battery capacity. The signature-aware policy beats them if it lowers the retrained deep attackers' detection AUC, not merely our probe's AUC, at no higher energy cost.
The key ablation is truncation level. If level-1 probes on (P, Q, t) match level-2, then ordering and area add nothing beyond channel increments and the geometric story fails.
Why not just…?
Why not just retrain a strong deep attacker on the defended data? This is the real threat, and we run it as the main baseline. It measures leakage well for one defense at a time. It is expensive across many defense, budget, and house combinations, and it does not say which structure leaks, which is what the defense design needs. If deep attackers and the signature probe disagree in ranking, the probe fails as an audit tool and we will report it that way. The defense-design claim is judged only against deep attackers, so it does not rest on the probe.
Why not just use classic hand-crafted event features such as step changes in P and Q and on-durations? These are essentially level-1 signature terms plus the time channel, so they are included as the level-1 ablation. The proposal is justified only if level-2 terms, which encode which channel moved first and the enclosed P–Q area, add measurable detection power under defense. That is the decisive check described below.
Why not just extend mutual-information analysis to the joint (P, Q) process? Exact mutual-information computation in Li et al. relies on small discrete state spaces. Joint, continuous, temporally correlated (P, Q) windows make it intractable. Plug-in estimates on marginals also miss ordering within a window. The level-2 versus scalar comparison under the Li et al. policy directly shows whether scalar-marginal guarantees miss leakage. Because real demand is not i.i.d., some residual leakage is expected even in P. We therefore attribute leakage to multichannel geometry only through the difference between the level-2 probe and the scalar and level-1 probes.
Decisive Checks
The first check is the gap. On UK-DALE, under the Li et al. and store-and-hide policies, we test whether the level-2 (P, Q, t) probe's AUC exceeds the level-1 and scalar-P probes for fridge, washing machine, and dishwasher by a margin outside bootstrap confidence intervals across held-out houses. The second check is the audit. The probe's defense ranking must correlate with retrained STAFN and Seq2Seq rankings at the agreement threshold stated in the Baselines section. The third check is the fix. At matched capacity and cost, the signature-aware battery must reduce retrained deep-attacker AUC relative to store-and-hide. If it only reduces our own probe's AUC, we will report that it overfits to the linear attacker.
If the first check fails, meaning scalar-optimal defenses leave no extra multichannel leakage, that negative result is itself useful. It would support the scalar modelling assumptions the battery literature makes.
References
- A Fundamental Bound on Performance of Non-Intrusive Load Monitoring with Application to Smart Meter Privacy
- Privacy-optimal strategies for smart metering systems with a rechargeable battery
- Smart Meter Privacy With Renewable Energy and an Energy Storage Device
- Privacy Preserving in Non-Intrusive Load Monitoring: A Differential Privacy Perspective
- Empirics on the expressiveness of Randomized Signature
- Dynamic Survival Analysis with Controlled Latent States
- An efficient representation of chronological events in medical texts
- Rough Transformers: Lightweight and Continuous Time Series Modelling through Signature Patching
- Convolutional Sequence to Sequence Non-intrusive Load Monitoring
- Sequence-to-point learning based on spatio-temporal attention fusion network for non-intrusive load monitoring
- CPDA: Class-Conditional Path Distribution Alignment for Unsupervised Time-Series Domain Adaptation