Catching Coordinated Small Data Injections in the Grid with Private Population-Level Distribution Tests

Smart-grid software that schedules household demand trusts the consumption forecasts each home reports. An attacker who controls many homes can nudge each report so slightly that no per-home anomaly detector notices, yet the combined effect still distorts prices and shifts costs onto honest neighbours. The strongest existing layered defense for this setting, Guarding the Grid, explicitly fails when injections change less than 0.1% of daily demand. It also requires homes to send detailed load profiles to the aggregator, a privacy cost its authors leave unresolved.

We propose adding one further safeguard that looks at the whole population rather than at individual homes. Each day, every home contributes a single weather-adjusted scalar, such as its normalised forecast residual. A differentially private two-sample test then checks whether the distribution of these scalars over recent days matches a trusted reference period with similar weather. Differential privacy is a guarantee that the test's output barely changes if any one home-day is removed. The test we reuse is the private continuous equivalence tester of Sheffet and Omer. In two sentences: per-home detectors miss small attacks because each home's signal is buried in its own noise, but a coordinated attack shifts the population distribution, and a population test pools evidence across thousands of home-days. Running that test privately on weather-adjusted residuals lets it sit alongside the existing detectors without exposing raw profiles and without raising alarms whenever a cloud front passes.

Core Claim

The claim is that the per-home layers and the population layer fail on different attacks, and that together they leave no profitable evasion region. Per-home detectors catch large, concentrated injections. The population test catches small, widespread ones.

Suppose a fraction $f$ of homes each shift their summary by $\delta$ relative to a per-home noise scale $\sigma$. The resulting distance between the population distributions is then of order $f\delta/\sigma$ for small shifts. This follows because a location shift changes the probability of each interval by roughly its density times $\delta$. By contrast, a per-home detector needs $\delta$ of order $\sigma$ before it can see anything.

The non-private term $\sqrt{k}/\alpha^2$ in Sheffet and Omer's bound says that $n$ samples resolve distances $\alpha \approx (\sqrt{k}/n)^{1/2}$. Here $k$ is the number of intervals over which distance is measured, $\alpha$ is the distance to detect, and $n$ is the number of home-days. The privacy terms add a cost that shrinks as $n$ grows. The detectable per-home shift therefore falls with population size, while the per-home threshold does not.

Two design choices make this usable. First, testing residuals from a weather-aware forecast, against a reference period matched on clear-sky index, separates attacks from legitimate weather shocks. That confusion is the main failure the file identifies for spectral-residual detectors. Second, choosing the home-day as the privacy unit keeps the tester univariate. That keeps us inside the regime where the Sheffet and Omer guarantees hold, rather than claiming a multivariate extension that does not exist.

Several things are deliberately out of scope. We do not place the test in the sub-second inverter control loop. Billing and day-ahead demand-response fraud operate on a daily timescale, so the tension between latency and sample size flagged in Gap 2 does not bind here. We do not claim to solve private testing over spatial fields. We also do not add noise to the power telemetry itself. Only the test statistic is privatised, so the Kirchhoff-law violations of Gap 3 do not arise.

Prior Work and Its Role

Guarding the Grid motivates the proposal and supplies the per-home layers. It is a three-stage pipeline:

  • a supervised convolutional network for known attack shapes;
  • a cluster-based spectral-residual model for novel ones;
  • isolation-forest search and double clustering to locate and repair corrupted time slots.

It documents both the sub-0.1% blind spot and the privacy exposure. We keep its pipeline intact and add a layer that targets exactly what it misses.

Sheffet and Omer supplies the method. It is a private equivalence tester for distributions on the real line. Bins are defined by sorted ranks rather than raw values, a random Bernoulli shift of each bin boundary stops one inserted point from cascading through all bins, and Poisson subsampling keeps stages independent. Its sample-complexity bound is the source of our scaling expectation. The file reports no empirical evaluation of it, so part of our contribution is the first measurement of its power on real telemetry.

Aliakbarpour et al. privatise discrete tests by averaging the statistic over all permutations of the sample. It is a natural alternative tester, but its guarantees are proven only for discrete alphabets. We therefore use it only after quantisation, as a comparison.

On the defense-in-depth side, two papers supply the evaluation lens rather than the method. Laszka et al. show for industrial control networks that combining diverse safeguards lowers risk by an order of magnitude compared with investing in one kind. Their model is static and contains no runtime statistical testing. Apruzzese et al. argue that defenses should be judged end to end against economically rational attackers, noting that only 20% of surveyed papers evaluate whole systems. We adopt their framing: the headline metric is the financial gain left to an attacker who knows every layer.

MixDefense is the closest precedent for pairing a statistical layer with a learned layer, split by perturbation size. It works on images, uses heuristic thresholds, and gives no privacy guarantee. We transfer its large-versus-small split to time series with a formal test.

For the weather adjustment we reuse forecasting machinery:

  • the clear-sky subtraction idea of Modified Auto Regressive Technique;
  • the spatiotemporal graph forecaster of Karimi et al., which reports a 120-minute mean absolute percentage error of 11.01%, against 18.98% for a temporal-convolution baseline.

Its residuals are the inputs we test.

Floating-point leakage is a constraint, because naive samplers can void the guarantee, as Ding et al. show for noisy-max selection. We therefore draw the tester's noise with a discrete sampler in the style of SampCert.

Research Program

We will build the population layer as a daily service run by the aggregator.

  1. Summary. Each home computes a small number of scalars per day. The primary scalar is the day's forecast residual normalised by that home's historical residual spread. A secondary scalar is the fraction of reported energy falling in peak-price hours, which targets attacks that move load in time while preserving daily totals.
  2. Reference matching. The aggregator keeps a reference pool of home-days believed clean, stratified by a daily clear-sky-index bin.
  3. Testing. For each summary, the private equivalence test compares the last $W$ days against the reference stratum matched on weather. The privacy budget is split across the few summaries and over a fixed horizon of repeated tests by standard composition.
  4. Alarm handling. An alarm hands control to the existing isolation stage of Guarding the Grid, which narrows down the responsible homes.

The analysis has three parts.

The first is the power curve of the private test as a function of compromised fraction, per-home shift, window length and privacy budget. We will compare it with the scaling predicted from the Sheffet and Omer bound.

The second is a coverage map over the plane of compromised fraction against per-home magnitude. It shows which layer catches which attack and whether any region evades both.

The third is the adaptive attacker. An attacker who knows all thresholds chooses injections to maximise bill reduction subject to evading every layer. We report the attacker's best achievable gain, which is the number that matters to the utility.

Baselines

The naive reference point is no defense. Guarding the Grid reports that injecting attacks into 5% of profiles yields a 17.8% bill reduction for the attacker. We reproduce this on the same Pecan Street setup: 168 Austin homes over three months, with pulse, scaling, ramping and random injections. That reproduction anchors the attacker-gain metric.

The strongest existing method is the full Guarding the Grid pipeline. It reports 94.42% precision and 0.19% false-positive rate for the supervised network, 86.53% recall for the spectral-residual model, 93.63% slot-isolation accuracy, and complete neutralisation of attacker gain for the attacks it detects. Its reported weakness is attacks below 0.1% of daily demand. We will re-run it on identical splits and identical injected attacks, extended to that small-magnitude regime. Every method receives the same training data, the same reference period, and the same false-alarm budget of 0.19% per day, matching the reported rate. Beating it means detecting coordinated sub-0.1% injections at that false-alarm rate, and lowering the adaptive attacker's best evasive gain, when our layer is added on top of the pipeline.

The power ceiling is a non-private two-sample test applied to the same weather-adjusted summaries. We use a Kolmogorov–Smirnov or Anderson–Darling test with an equal false-alarm budget. It shows how much detection power privacy costs.

The quantise-then-privatise baseline bins the same summaries. It then applies a Laplace-noised chi-square test, or the permutation-smoothed tester of Aliakbarpour et al., at the same privacy budget. Bin counts are tuned on the reference period only.

The aggregate-total check compares daily summed reported energy with feeder-level totals.

Finally, two ablations isolate the key ingredient. The first runs the private test on raw reported consumption instead of weather-adjusted residuals. The second uses residuals with a reference that is not weather-matched. These ablations show whether the weather adjustment is what separates attacks from cloud fronts.

Why not just…?

Why not just retrain the per-home detectors on small attacks? A per-home classifier sees only one home's series. When the injection is well below that home's natural forecast error, no classifier can separate attacked from clean days in that home, however it is trained. We will test this directly by retraining the supervised network and spectral-residual model with sub-0.1% injections included. We expect detection near the false-alarm rate. If retraining instead succeeds at the 0.19% false-alarm budget, the population layer loses its main motivation and we will report that.

Why not just check summed reported energy against the feeder meter? This is a real threat. Against attacks that change daily totals it may work as well as our test, given a trusted, well-synchronised feeder meter. Our case rests on two conditions. The first is total-preserving attacks that shift reported load into cheaper hours, which a totals check cannot see by construction. The second is deployments where feeder metering is coarse or itself untrusted. The experiments include both total-changing and total-preserving attack families, so the comparison will show exactly where the simpler check suffices.

Why not just run a non-private two-sample test, or protect inputs with secure aggregation? A one-bit alarm leaks little per day. However, the aggregator runs the test daily, logs statistics, and passes alarms to an isolation stage. Over months, these repeated outputs add up to a meaningful leakage channel, and secure aggregation alone does not bound it. Differential privacy does bound it under composition. Whether that bound is worth paying for is an empirical question, settled by the gap between the private test and the non-private ceiling. If the gap is small at a modest privacy budget, privacy is nearly free. If it is large, we will say that the private version is only justified where regulation demands it.

Why not just quantise and use discrete private tests? The research file argues that narrow bins inflate noise and wide bins erase small shifts. The rank-based bins of the continuous tester adapt to the data without spending privacy on choosing boundaries. The quantised baseline, tuned fairly, will show whether this advantage appears at realistic sample sizes.

Proposed Experiments

Pecan Street. The primary experiment uses Pecan Street, as configured in Guarding the Grid. With the home-day as the privacy unit, this gives about 15,000 home-days. That number is small for the sample-complexity bound, so we will also report power on a resampled synthetic population of up to several thousand homes. The synthetic homes are drawn from fitted per-home residual models, which lets us check the predicted improvement with population size. Attacks follow the four published families. We add a coordinated family in which a fraction $f$ of homes, from 1% to 30%, each inject between 0.01% and 1% of daily demand, in both total-changing and total-preserving variants. Metrics are:

  • detection probability at the fixed false-alarm rate;
  • detection delay in days;
  • the adaptive attacker's maximum evasive bill reduction, compared with the reported 17.8% without defense;
  • cost to honest neighbours.

False alarms from weather. The second experiment tests the weather confounder with no attacks injected. We run the population test on residuals from the spatiotemporal graph forecaster over the 316-site photovoltaic dataset of Karimi et al., and on photovoltaic-equipped Pecan Street homes. We measure the false-alarm rate on days with large irradiance swings, for the raw, unmatched-residual and matched-residual variants.

Decisive checks. Three outcomes would settle the proposal. First, at the fixed false-alarm budget, the combined pipeline must detect coordinated sub-0.1% attacks that the reproduced Guarding the Grid pipeline misses. Second, the failure sets of the per-home and population layers must be largely disjoint on the coverage map, which is the orthogonality that layered defense depends on. Third, weather matching must keep false alarms on high-variability days near the nominal rate, while the raw-value ablation does not.

If the first check fails at realistic population sizes, or the feeder-totals check matches our layer on total-preserving attacks, the idea does not hold in its current form.