acceptodds
Under review as a conference paper at ICLR 2027

Learning Defensive Policies Against Diverse Inference Attacks for Smart Meter Privacy

Abstract

Smart meter (SM) data provides fine-grained visibility into household energy consumption, but also exposes users to privacy risks. Inference attacks, known as non-intrusive load monitoring (NILM), can perform appliance-level inference from aggregate signals and recover sensitive behavioral patterns. In practice, attacker models are unknown and heterogeneous, making robust defense challenging. We formulate SM privacy protection as a black-box inference defense problem, aiming to reduce the recoverability of appliance-level information while generalizing across diverse and unseen attackers. We propose a proxy-guided hierarchical reinforcement learning framework that learns battery-based load-shaping policies to inject realistic but misleading appliance-level signatures into the aggregate signal, thereby disrupting the structured patterns exploited by NILM. A self-supervised aggregate-structure privacy probe provides a reconstruction-error-based surrogate reward for disrupting recoverable load structure, while a signature library makes the perturbations appliance-relevant and physically realizable through battery control. We provide theoretical rationale showing that proxy-guided optimization improves inference robustness under attacker diversity. Experiments on real-world datasets UK-DALE and REDD demonstrate strong cross-model and cross-appliance generalization. Across six unseen NILM attackers, covering four appliances on UK-DALE and five on REDD, our proposed defense increases average appliance-level RMSE by 107% and 166%, respectively, while reducing F1 score by 79% and 80%.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
smart meter privacy, non-intrusive load monitoring, privacy signal, black-box attack, reinforcement learning, adaptive attacker

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3vqBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a battery-based defense against appliance inference from smart-meter data. A reinforcement learning policy selects appliance-like patterns to inject into aggregate power readings, guided by a self-supervised reconstruction model. Experiments on UK-DALE and REDD evaluate its effectiveness against multiple inference models, including adaptive attackers.

Critical strengths
  1. The paper addresses a practical privacy problem while respecting physical battery constraints.

  2. The experiments cover diverse attackers and two datasets, including adaptive attackers.

Critical weaknesses
  1. The experiments provide empirical support for the proxy, but its reliability and limitations remain unclear. The paper does not sufficiently explain when aggregate reconstruction error is a useful indicator of appliance-level privacy.

  2. The methodological advance over existing approaches is not clearly established. The paper should better distinguish the contribution of the proposed combination from that of its established components.

  3. The theoretical conclusions are conditional on alignment between the proxy and attack objectives. The paper does not establish whether this condition holds for the proposed framework, which limits the explanatory value of the theory.

  4. The component ablations do not fully rule out differences in perturbation strength as an explanation for the gains. Matching battery constraints alone does not establish that the methods use comparable perturbations.

  5. The REDD split description appears inconsistent with the main results and needs clarification. The paper also does not clearly explain how differences in final battery charge are accounted for in the cost comparison.

Questions for authors
  1. The experiments show that the proxy can be useful, but its limitations are less clear. When might a higher aggregate reconstruction error fail to indicate better appliance-level privacy?

  2. The method combines a reconstruction model, reinforcement learning, and an appliance signature library. Could you clarify the main methodological advance over prior battery-based privacy defenses?

  3. The theory assumes alignment between the proxy and attack objectives. What evidence suggests that this assumption holds in your experiments?

  4. Could differences in perturbation strength partly explain the gains? The methods share battery limits, but it is unclear whether they actually introduce comparable perturbations.

  5. Does the defense make appliance usage inherently more ambiguous, or mainly shift the input away from the attackers’ training distribution? What do the retraining results tell us about this distinction?

  6. Could you clarify the REDD training and test dates, which appear inconsistent across sections? Also, does the cost comparison account for differences in the battery’s final charge?

  7. What is the attacker assumed to know about the defense? In particular, do the adaptive-attacker experiments assume knowledge of the signature library and control policy, or only access to defended training data?

1 comment

We thank reviewer 3vqB.

W1 & Q1

Where the proxy holds and where it fails. Holds: along 100 policy updates, changes in probe reward correlate positively with the error of all six held-out attackers (pooled ρ = 0.43 / 0.45 for RMSE / F1). The proxy is informative when perturbations change aggregate features used for appliance inference; changes unrelated to appliance events are not expected to reduce privacy. Fails: privacy is measured as a real attacker's degradation, so an attacker that already fails on undefended data has little left to lose. The lowest correlations are such pairs: CO / FHMM model discrete on/off states and miss the near-continuous fridge-freezer even without defense (RMSE 79 / 77 W vs. 73 W for a constant), and DAE barely detects the microwave (raw F1 0.31). On REDD, the only 6 / 30 RMSE decreases are CO / FHMM on Light and Sockets (raw F1 0.01–0.40), with no significant F1 rise. Revision: new "Scope of the proxy" paragraph in Sec. 4.

W2 & Q2

Main advance: label-free probe reward and signature actions. Prior battery defenses (DDQL, TCN, PLS) act per minute in a flat action space with smoothness, MI or load-leveling rewards. We change both: (i) reward: reconstruction error of a fixed self-supervised probe, needing no labels or attacker; (ii) action: an appliance signature executed by a hierarchical controller under battery constraints. Each part contributes (Table 3, RMSE / F1): flat + conventional 63.8 / 47.1%, signature + conventional 80.5 / 59.6%, flat + probe 87.6 / 64.8%, both 107.1 / 79.3%. Against PLS-DDPG directly (App. F.3, Table 19) we are competitive (F1: 5 better, 2 worse, 17 n.s.) with 30% less throughput and 3.7× less training time. Revision: Sec. 1 lists (i)–(ii) as the contribution, building on a standard battery model and PPO.

W3 & Q3

Empirical evidence for A1. We examine the alignment suggested by A1 along policy updates: 21 training checkpoints × 5 seeds, six fixed held-out attackers, 24 cases. Spearman ρ between Δprobe reward and Δattacker loss over 100 updates is positive for all six attackers: for RMSE 0.33–0.44 on the four neural attackers and 0.19 / 0.18 on CO / FHMM; for F1 0.37–0.46 except DAE (0.16); pooled 0.43 / 0.45 (p < 10⁻⁵). Over the first 20 episodes RMSE / F1 degradation rises from 44.9% / 33.1% to 102.6% / 76.9%. Revision: Sec. 4 states A1 as a sufficient condition and adds the ρ table; App. B.6 is rewritten as "Proxy–Attacker Alignment along Policy Updates"; the abstract and contribution (3) are narrowed to match.

W4 & Q4

Gains are not explained by larger perturbations. All variants share the same feasible set; within it, how much of the budget to use is part of the learned policy, so equal constraints are the operationally fair comparison. Our perturbation is also structured against NILM attackers rather than made as large as possible; to rule out perturbation strength, we report actual throughput and perturbation. Ours reaches 113.9% / 78.2% (RMSE / F1) at 5.63 kWh/day and 472 W. Random library signatures at the same throughput reach only 74.6% / 38.0% despite a 25% larger perturbation (p ≤ 0.002). At matched throughput and perturbation (5.73 kWh/day, 485 W), the perturbation-energy reward reaches 97.3% / 71.3%, 16.6 / 6.9 points below ours, and its final policy needs a 28% larger perturbation for a similar effect (p = 0.73 / 0.82); PLS-DDPG uses an 83% larger one with a lower F1 reduction. Revision: new "Perturbation magnitude" paragraph and Table 10 in Sec. 6; App. B.7.

W5 & Q6

Dates fixed; cost compared at equal terminal energy. (1) REDD House 2: train 2011-04-19 – 04-28, main test 04-29, extra held-out 04-30; UK-DALE House 1: train 2013-03-19 – 03-28, test 03-18; no overlap. (2) Settling each method's end-of-day SoC back to 50% at the off-peak tariff, ours costs −£0.38/day (−£0.19 with recharge degradation), the lowest of all methods. Revision: App. E.1 corrected; new App. A.3 and a "Total (adj.)" column in Table 2.

Q5

Protection persists after attacker retraining. Retraining on labelled defended data lets attackers adapt to the defended distribution. Mean F1 is 0.723 unprotected, 0.161 fixed, 0.284 (scratch) and 0.265 (fine-tuned): retraining recovers only 0.12 F1, 12–32% of the loss per architecture, and no architecture exceeds 0.325 mean F1. The defense stays effective even after attackers adapt to labelled defended data, consistent with signature decoys that are hard to tell from genuine appliance events. Revision: Table 9 now reports per-architecture results and recovery rates.

Q7

Attackers never access the library or policy. Fixed attackers (Tables 1–2) have clean training data only. Adaptive attackers (Table 9) additionally train on the deployed defender's output on the 10 training days with appliance labels. Neither has the signature library or the policy parameters; they see them only through labelled defended data. Revision: threat model added to Sec. 3.

Reviewer_dRJ1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses battery charging and discharging to insert misleading appliance-like patterns into smart-meter readings. A reinforcement learning policy selects these patterns using a reconstruction network as a privacy reward, without accessing the target attacker. Experiments on UK-DALE and REDD evaluate six attackers, existing defenses, component contributions, and attacker retraining.

Critical strengths
  1. The evaluation covers several attacker architectures and includes retraining on defended signals. The remaining reduction in F1 after retraining provides stronger evidence than testing only fixed attackers.

  2. The controlled experiment in Table 3 separates the contributions of the signature library and the probe reward. Both contribute to the observed improvement, supporting the practical value of combining them.

Critical weaknesses
  1. The REDD split is inconsistent. Table 8 identifies April 29 as the main test day, while Appendix E.1 includes April 29 in training and reserves April 30 for testing.

  2. The significance summary conflicts with the detailed results. Table 13 shows RMSE decreasing in six REDD cases, so Table 1 cannot report significant RMSE increases in 26 of 30 cases. Significant changes that improve attacker performance should not count as privacy gains.

  3. Each test day starts with the battery at 50% charge, regardless of how much charge remains at the end of the previous day. In practice, the next day must start with the remaining charge, which changes how much the battery can charge or discharge. The separate daily tests therefore do not show whether the defense can maintain its effectiveness over consecutive days.

Questions for authors
  1. If the controller requests more battery discharge than the household currently needs, does the executor reduce the discharge to match demand? If surplus electricity is instead exported to the grid, how is it accounted for in the reported cost savings?

  2. In the REDD experiments, do you directly use the controller trained on UK-DALE, or train a new controller on REDD? Please also clarify whether the privacy probe is retrained or whether both the controller and probe are reused unchanged.

3 comments

We thank reviewer dRJ1 for the careful reading.

W1

REDD split. April 29 is the main REDD test day and was never used for training; Appendix E.1 listed the training range incorrectly, and we have fixed it to match Tables 1 and 8. All reported REDD results use the split below, in which the last two of the 10 training days serve as validation (1-min resolution, local time):

DatasetTrain (10 days)Main testAdditional held-out
REDD House 22011-04-19 00:00 – 04-28 23:592011-04-292011-04-30; H1, H3, H4
UK-DALE House 12013-03-19 00:00 – 03-28 23:592013-03-18other seasons; H4

W2

Significance direction. We agree: a two-sided test also flags significant changes that help the attacker, and these should not count as privacy gains. Table 1 now counts only Holm-significant changes in the privacy direction (RMSE increase, F1 decrease):

DatasetVersionRMSEF1
UK-DALE (24)Manuscript2224
UK-DALE (24)Revised2224
REDD (30)Manuscript2619
REDD (30)Revised2219

Only REDD RMSE changes, from 26/30 to 22/30. All six REDD RMSE decreases are CO / FHMM on Light, Sockets 1 and Sockets 2, where these attackers already fail without defense (raw F1 0.01–0.40), so there is little appliance information for the defense to remove. Event detection improves in none of them: F1 falls in 5 of 6 (significantly for CO–Light, FHMM–Light, FHMM–Sockets 2, p ≤ 0.02), and the one rise (+0.03) is not significant (p = 0.43).

PositionManuscriptRevised
Table 1, REDD RMSE Sig.26/3022/30 (privacy direction)
Holm table and testing protocol (App. F.2)two-sided countsprivacy direction only
Sec. 6, l. 349"consistently raise NILM reconstruction error""raise RMSE in all 24 UK-DALE and 24 of 30 REDD cases"

W3

Consecutive days (new App. A.4, Table 12). The defense maintains its effectiveness when the next day starts from the remaining charge. The current SoC is part of the defender's observation, so the policy acts on the charge actually left rather than on a fixed 50% start; training therefore needs no terminal-SoC constraint.

Setup. The trained defenders (5 seeds, no retraining) run 5 consecutive days with SoC carried over and 95% charge / discharge efficiency, on four 5-day blocks of held-out UK-DALE House 1 days, against the six attacker architectures on the four appliances of Table 1 (23 cases; FHMM–fridge-freezer excluded, as its output does not depend on the input). As a reference, a forced 2 kW recharge brings SoC back to 50% by 23:59 each day; p-values compare the two policies (paired t-test).

SettingRMSE inc.F1 red.kWh/dayEnd-of-day SoCCost (£/day)
Continuous, original policy98.5%63.3%7.000.150.17
Forced recovery to 50% each evening97.2% (p = 0.64)64.8% (p = 0.41)9.250.500.65 (p < 0.001)

Result. After day 1 the SoC settles into a stable daily cycle, and protection persists at £0.17/day (7% of the household bill). Forcing SoC back to 50% each evening does not change privacy, but needs 2.3 kWh/day more throughput and 3.8× the cost.

Q1

(a) Discharge is clipped to demand. Yes. At every minute the executor sets the executed discharge to the smaller of the requested discharge and the current household load (P_user(t) + P_batt(t) ≥ 0, now explicit in App. A.2), so the meter reading never becomes negative. SoC, cost and the throughput budget are updated with the executed power, not the requested one.

(b) Cost accounting. No electricity is exported or sold back to the grid (App. A.1), so the reported savings contain no feed-in revenue: they come only from shifting household consumption in time under the time-of-use tariff. A non-negative reading also keeps the masked load within the range of genuine household consumption. Export tariffs are utility-specific, and we add export-enabled operation as future work (Sec. 7).

Q2

REDD controller and probe. We train a new controller and a new probe on REDD; neither is reused from UK-DALE. Both use only the 10 REDD training days and the aggregate signal (no appliance labels). Only the signature library is shared: it stores appliance power trajectories, not dataset-specific statistics, and is reused unchanged. Within each dataset, the held-out-day and cross-household results (Tables 7–8) reuse that dataset's controller and probe without retraining. Appendix E.1 now states this explicitly.

I thank the authors for their responses. The corrected REDD split and significance counts, additional continuous-operation experiments, and clarifications on discharge constraints and REDD training substantially address my concerns. The revisions strengthen the empirical evidence. I will raise my rating from 2 to 3.

We sincerely thank reviewer dRJ1 for the careful review and for raising the score.

We are glad that the corrected REDD split description, the privacy-direction significance counts, the continuous-operation experiments and the clarifications on discharge and REDD training addressed the concerns.

All of these changes are included in the revised paper.

Reviewer_c1oLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies smart-meter privacy against unknown non-intrusive load monitoring (NILM) attackers. It proposes a battery-based load-shaping defense combining a fixed, self-supervised aggregate reconstruction probe with a library of appliance signatures. A PPO-trained high-level policy selects signatures and replay directions, while a deterministic executor enforces battery constraints. The training reward measures the increase in probe reconstruction error caused by the defense. Experiments on UK-DALE and REDD evaluate transfer across six NILM attackers, compare against existing battery-based defenses, and include component ablations, battery sensitivity, household and seasonal generalization, and adaptive-attacker evaluations. The theoretical analysis gives a local sufficient condition under which improving the proxy objective increases expected attacker inference error.

Critical strengths
  1. The defender does not require access to the evaluation attackers, and the deployed policy uses aggregate measurements and battery state rather than appliance labels. The component isolation in Table 3 is particularly useful: it supports separate contributions from the probe reward and signature-based action space, rather than attributing all improvements to an undifferentiated framework.
  2. The paper evaluates heterogeneous NILM methods, compares against multiple existing defenses under common battery constraints, and studies different probe architectures and battery configurations. It also includes retraining and fine-tuning attackers, which is substantially more informative than evaluating only frozen attackers. These experiments provide promising evidence that the approach can transfer beyond its training probe.
Critical weaknesses
  1. Condition A1 in Appendix B assumes alignment between the proxy-risk gradient and the expected attacker-risk gradient in the policy-induced metric. The subsequent improvement result follows from this assumption and local smoothness, but neither task commonality nor aggregate identity reconstruction establishes A1. The final transfer results are consistent with useful alignment, but do not directly validate the local gradient condition. Moreover, a large reconstruction error against the clean aggregate can reflect perturbation magnitude rather than certify departure from an NILM-relevant region. The stronger claims in Section 4 therefore exceed what the analysis establishes.

  2. Appendix E.1 places REDD training data between April 19 and April 29 and identifies April 30 as the held-out test day, whereas Tables 1 and 8 identify April 29 as the main evaluation day. This needs reconciliation to exclude training–evaluation overlap. The adaptive-attacker results are encouraging, but their training duration, defended-data construction, validation procedure, and policy/seed separation are not sufficiently specified to assess whether the attackers were adequately optimized. In addition, improvements over the strongest baseline are relatively modest, while significance counts primarily compare each defense with no defense rather than directly testing the proposed method against that baseline.

Questions for authors
  1. Can you directly measure proxy–attacker alignment, or the relationship between changes in proxy reward and changes in held-out attacker loss, across feasible policy updates? Does the proposed reward remain more predictive than a simple perturbation-energy objective when actual battery throughput and perturbation magnitude are matched?
  2. What are the exact REDD training, validation, and test timestamps? Please reconcile the April 29 main evaluation in Table 8 with Appendix E.1. For adaptive attackers, please provide the defended-data generation protocol, training budgets, validation and model-selection criteria, and separation between training and evaluation days and defense seeds.
  3. Do the privacy and cost results persist under continuous multi-day operation without resetting SoC to 50% each day, with realistic battery efficiencies and a terminal-energy constraint? How is the stated no-export setting enforced when a discharging signature exceeds the current household load?
2 comments

We thank reviewer c1oL for the careful reading.

W1 & Q1

W1. A1 as a sufficient condition. A1 is a sufficient condition on the attacker family: under it, Propositions 1–2 guarantee that proxy ascent increases attacker risk. We narrowed Sec. 4 accordingly: it no longer states that the probe reward "certifies" privacy or "provably increases" attacker risk, only that it does so under A1.

W1/Q1. Proxy–attacker alignment (Sec. 4; App. B.6). To answer Q1, we directly measure how changes in proxy reward relate to changes in held-out attacker loss across feasible policy updates: 21 training checkpoints × 5 seeds, attacked on the test day by the six fixed held-out attackers on four appliances (100 updates, 24 cases).

Spearman ρPooledBERT4NILMELECTRIcityDAES2PCOFHMM
RMSE0.430.330.370.350.440.190.18
F10.450.390.460.160.370.420.42

Bold: p < 0.05. The correlation is positive for every attacker; the low entries are pairs that already fail without defense.

W1/Q1. Perturbation magnitude (Sec. 6, Table 10; App. B.7). Our perturbation is structured against real NILM attackers, not made as large as possible. We compare policies at similar throughput and perturbation magnitude on the same 24 cases: a perturbation-energy reward (same defender), random library signatures under a daily budget, and PLS-DDPG.

PolicyThroughput (kWh/day)Perturbation (W)RMSE / F1 degradation
Ours5.63472113.9% / 78.2%
Random, 5.6 kWh budget5.6058974.6% / 38.0%
Random, 8.0 kWh budget8.00666109.7% / 60.5%
Perturbation-energy, matched checkpoints5.7348597.3% / 71.3%
Perturbation-energy, final policy7.24605110.9% / 77.6%
PLS-DDPG7.98862114.8% / 70.5%

Learned policies are retrained here and RMSE is summed over cases. At equal throughput, ours beats random signatures by 39.3 / 40.2 points (p ≤ 0.002) with a smaller perturbation. At approximately matched throughput and perturbation (5.73 vs. 5.63 kWh/day; 485 vs. 472 W), the perturbation-energy reward reaches 97.3% / 71.3% vs. our 113.9% / 78.2%, supporting the probe reward beyond perturbation magnitude; its final policy needs a 28% larger perturbation to match us (p = 0.73 / 0.82), and PLS-DDPG uses an 83% larger one and still trails in F1.

W2 & Q2

W2/Q2. REDD dates. The typo in App. E.1 is fixed; there is no overlap. REDD House 2: train 2011-04-19 – 04-26, validation 04-27 – 04-28, main test 04-29; extra held-out H2 04-30, H1 04-26, H3 05-25, H4 05-01. UK-DALE House 1: train 2013-03-19 – 03-26, validation 03-27 – 03-28, test 03-18.

W2/Q2. Adaptive attackers (App. F.4). The deployed defender (seed 42) is rolled out on the 10 training days and, separately, on the test day. Attackers train on the first 80% and validate on the last 20% of the training days in time, keeping the lowest-validation-loss checkpoint (early stopping), with each model's original budget (BERT4NILM 20, ELECTRIcity 10 + 90, DAE / S2P ≤ 40 epochs); scratch retraining and fine-tuning differ only in initialization. All are evaluated on the disjoint defended test day; training and test data come from the same deployed defender (seed 42), as an attacker observing that system would collect them, and no other seed is used. Every architecture improves over its fixed version and both initializations converge to within 0.04 F1, yet retraining regains only 12–32% of the lost F1 (Table 9).

W2. Direct test against PLS-DDPG (App. F.3, Table 19). Paired t-test per case, Holm-corrected: F1 — ours better in 5, no significant difference in 17, worse in 2; RMSE — 0 / 23 / 1. Our method is competitive (Table 2: 107.1% / 79.3% vs. 98.1% / 75.8%) while using 30% less throughput and 3.7× less training time.

Q3

Continuous operation (App. A.4, Table 12). SoC is in the Manager's observation, so training needs no terminal-SoC constraint. Trained defenders (5 seeds, no retraining) run 5 consecutive days without SoC reset at 95% charge / discharge efficiency, on four 5-day blocks of held-out UK-DALE House 1 days. Protection persists (98.5% / 63.3% RMSE / F1) at £0.17/day; forcing SoC back to 50% each evening gives 97.2% / 64.8% (p = 0.64 / 0.41) but costs £0.65/day.

Cost under equal terminal energy (Table 2; App. A.3). This adjustment only normalizes the single-day cost comparison; continuous operation above carries the actual SoC between days. Settling each method's end-of-day SoC back to 50% at the off-peak tariff, ours costs −£0.38/day (−£0.19 with recharge degradation), the lowest of all methods and the only one below no defense with recharge degradation (PLS-DDPG −£0.03 / +£0.18).

No export. Discharge is truncated to the current load, so nothing is exported (now explicit in App. A.2).

AuthorsOpenReview ↗

Brief clarification of the alignment test results

We thank the reviewer again.

As a brief clarification of the alignment test (App. B.6): the three lower entries (CO / FHMM RMSE 0.19 / 0.18, DAE F1 0.16) are attacker–appliance pairs that already fail on undefended data, so they have little appliance information left to lose.

CO / FHMM predict the near-continuous fridge-freezer worse than a constant at its mean (RMSE 79 / 77 W vs. 73 W), and DAE barely detects the microwave (raw F1 0.31).

For event detection, the correlation is positive in all 24 attacker–appliance cases, and the CO / FHMM F1 entries remain significant (0.42). We would be glad to clarify anything further before the discussion period ends.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.