Fooling Every Extractor at Once: Multi-Backbone Latent Perturbations for Smart Meter Privacy

A smart meter records a home's total power draw many times per second. Non-intrusive load monitoring (NILM) algorithms can take that single aggregate stream and split it back into per-appliance traces, exposing when residents sleep, shower, cook, or run medical equipment. One line of defense repurposes adversarial attacks: the meter itself adds a small perturbation to the reported power stream, computed from the gradients of a surrogate NILM network, while keeping the perturbation sum exactly zero over each billing cycle so the total energy billed is unchanged. The strongest existing scheme of this type, ADV-NILM, drives a disaggregator's mean absolute error on REDD from 24.15 W to 174.91 W with zero billing error. But it computes its perturbation against one surrogate network, and this is where a second research direction bites.

Ensemble feature extraction shows that differently structured networks — say a densely connected CNN, a residual network, and a transformer — learn internal representations that live in essentially different coordinate spaces, even when their final predictions are highly correlated. Feature-level fusion of several backbones is more robust than averaging their outputs precisely because each backbone covers the others' blind spots. That robustness is a gift to an attacker who fuses features from several architectures before disaggregating, and a hole in defenses that target a single architecture's gradients. This proposal builds and studies both sides of that contest: an ensemble-feature disaggregator as a realistic worst-case attacker, and a defense that synthesizes perturbations against a fused multi-backbone latent space rather than any single model.

Core Claim

The claim is two-sided. First, existing single-surrogate adversarial defenses degrade substantially against a disaggregator that concatenates latent features from architecturally diverse backbones (a dilated convolutional network, a sequence-to-sequence CNN, and a stationarization-aware transformer) before a prediction head, because the perturbation only corrupts the coordinate subspace its surrogate shares with each backbone. Second, and constructively, a defender that trains its own ensemble of diverse frozen backbones — cheaply, on public data, with no white-box knowledge of the attacker — and optimizes a single zero-sum perturbation to maximally disrupt the concatenated latent space can remain effective across attacker families, including statistics-based transformers that point-wise defenses miss. We name this defense FuseGuard. The mechanism has one equation worth writing: the meter solves

$$\max_{\delta}\; \sum_{k=1}^{K} \big\| f_k(x+\delta) - f_k(x) \big\|_2^2 \quad \text{s.t.} \quad \sum_t \delta_t = 0,\; |\delta_t| \le \epsilon,$$

where $f_k$ is the $k$-th backbone's latent map of the reported window $x$, $\delta$ is the injected perturbation, and the constraints preserve billing exactly and bound the instantaneous distortion. This differs from averaging the loss gradients of several models: the objective is measured in representation space before any prediction head, targeting features that downstream appliance classifiers actually consume.

Research Program

Three connected workstreams. The first is a threat-model audit. We build ensemble-feature disaggregators in the style validated elsewhere: frozen backbones with contrasting inductive biases, latents concatenated and passed to a trained head, following the fusion recipe of Multi-EPL, which showed concatenating independent extractors builds a more discriminative latent space, and the evidence from Penatti et al. that distinct topologies capture complementary coordinates of the data distribution. We include backbone diversity along two axes the literature says matter: architecture (convolutional, recurrent, attention) and optimizer, since DAMSL showed backbones trained under Adam versus SGD produce complementary coordinates for downstream learners. We also include a stationarized transformer attacker in the spirit of NILMFormer, which normalizes each window by its mean and standard deviation before attention; because a zero-sum perturbation preserves the window mean exactly but not its variance, this attacker is the sharpest test of whether point-wise defenses survive. We then run ADV-NILM's published single-surrogate perturbations against this ensemble and measure how much of its reported degradation is lost.

The second workstream is FuseGuard itself. The defender trains its own $K$ backbones on public datasets (REDD and UK-DALE are open and widely used), freezes them, and solves the constrained objective above with a projected optimizer in the Jaco-Adam lineage. Because the defender never sees the attacker's model, robustness comes from diversity of the surrogate ensemble, the same principle SwiftTheft used on the opposite side — aggregating reference models to cancel any single architecture's bias, achieving near-complete elimination of single-model bias in its vision setting — transplanted from model theft to load privacy. A practical wrinkle from the ensemble literature: Multi-EPL showed concatenation gains plateau and dimensionality inflates overfitting, so we compress the fused latent space with a learned projection before optimizing, which also keeps the meter's compute budget realistic for edge hardware.

The third workstream is analysis. Farokhi's Cramér–Rao bound shows NILM estimation error is lower-bounded via the cross-correlation of appliance load derivatives — ramp rates. We will measure whether FuseGuard's perturbations, discovered purely by gradient ascent in latent space, end up decorrelating exactly those derivative features, connecting the learned defense to the information-theoretic account of what makes disaggregation hard.

Baselines

The do-nothing reference point is releasing the clean aggregate stream. Under it, ADV-NILM reports a disaggregation MAE of 24.15 W on REDD for the unperturbed case, which sets the floor for attacker performance; our ensemble attacker should be at least this good on clean data, or the threat model is not credible. The strongest existing defense baseline is ADV-NILM itself (He et al.), whose headline numbers — REDD MAE rising from 24.15 W to 174.91 W, signal aggregate error from 12.55% to 695%, billing error exactly 0 — are the mark any new defense must beat while preserving the same zero billing error. It will be run with its own tuning budget on our identical data splits, and evaluated against both single-model and ensemble attackers, since its single-surrogate design is exactly the limitation we claim to fix. A second defense baseline is the GAP releaser of Shateri et al., which reduced an occupancy adversary's accuracy from above 80% to roughly 50% on ECO; on the ECO occupancy task we will run both defenses under matched distortion budgets. A third is Laplace noise at matched $\ell_\infty$ magnitude, the naive reference for any perturbation scheme. On the attacker side, the strongest competitors are AED (Du et al.), which reports MAE of 17.68 W on REDD and 9.15 W on UK-DALE with its multi-discriminator feature learning, and NILMFormer, which reports an average 15% MAE improvement over BERT4NILM and UNet-NILM across UK-DALE, REFIT, and EDF benchmarks; our ensemble attacker will be tuned to sit at or above these on clean data before any defense is applied. Finally, the key ablation isolates the proposal's ingredient: a same-budget defense that averages the output-space loss gradients of the same $K$ backbones (multi-model gradient ascent, no latent fusion). If that matches FuseGuard, the latent-space claim is falsified and we will say so.

Why not just…?

Why not just add differential-privacy noise? Three documented failures. Random DP noise breaks the billing total, since the perturbations no longer sum to zero, whereas ADV-NILM demonstrated exactly zero billing error under its constraint; the DP-perspective paper of Nassereldin et al. proved that because appliance switching events are over 98% sparse, compressive-sensing reconstruction survives DP noise injection; and in the power-system setting Rigos et al. showed naive Laplace noise largely destroys physical AC feasibility, requiring elaborate post-processing. None of this makes DP useless, but it does not dominate the constrained-adversarial approach on this task.

Why not just average the gradients of several models instead of fusing latents? This is the most serious alternative, and we treat it as the primary ablation rather than arguing it away. The evidence for expecting latent fusion to win is indirect but consistent: Penatti et al. showed correlated prediction errors hide uncorrelated latent geometries, so output-level objectives average away the very diversity that helps; SwiftTheft found output-level scoring failed to eliminate single-architecture bias where concatenated latent features succeeded; and AED's feature-level discrimination decoupled overlapping appliance signatures that prediction-level processing could not. If the ablation shows gradient averaging matches FuseGuard across all attacker families, we will report the simpler method as the recommendation — the ensemble audit workstream stands on its own either way.

Why not just lower the sampling frequency or install a battery? Coarse sampling destroys the sub-minute telemetry that local energy communities and voltage forecasting need, and NILMFormer demonstrates disaggregation still works at 10- and 30-minute sampling on EDF data with a 51% error reduction over regression baselines, so coarsening does not even buy safety. On the hardware side, the adversarial hypothesis-testing analysis of Fan et al. indicates blinding an optimal adversary requires roughly 4000–5000 W of battery or renewable capacity, economically prohibitive for most households, and even with solar influx disaggregation remains viable, as Zhang et al. showed. Software defenses cost nothing in hardware and preserve the aggregate exactly.

Proposed Experiments

All experiments run on REDD, UK-DALE, and REFIT for appliance-level disaggregation (fridge, dishwasher, microwave, kettle, washing machine), and on ECO for occupancy, the private attribute with ground truth. Metrics are MAE in watts, signal aggregate error, appliance-level F1, occupancy classification accuracy, and billing error, the last of which must be identically zero for any acceptable defense. The decisive checks are four. First, the audit: ADV-NILM perturbations transferred from a single surrogate lose at least half of their reported MAE inflation against the ensemble-feature attacker — if they do not, the threat motivating FuseGuard weakens and the proposal pivots to reporting that robustness result. Second, the head-to-head: FuseGuard versus ADV-NILM under identical distortion budgets and identical attacker ensembles, where success means holding perturbed MAE above the clean-data MAE of the strongest single attacker (17.68 W on REDD per AED) against every attacker family, including the stationarized transformer. Third, the ablation against multi-model gradient averaging, the honest test of whether representation-space objectives matter. Fourth, the analysis check: correlation between FuseGuard's perturbation-induced latent disruption and the decorrelation of load-derivative cross-correlations predicted by the Fisher-information bound, which would give the learned defense a theoretical footing. Runtime matters for an edge device, so we report the meter's per-window solve time and compare against SwiftTheft's reference-ensemble precedent of one-pass efficiency. No reported result found on Lacuna for any prior defense evaluated against a multi-architecture feature-fused disaggregator — that absence is the gap this proposal fills.