acceptodds
Under review as a conference paper at ICLR 2027

PEGAS: Security Assessment of Power Grid Agents through Physics-informed Evolutionary Graph Adversarial Search

Abstract

Reinforcement Learning (RL) agents and neural solvers are becoming the standard for safe & reliable power grid control. Standardized benchmarks and challenges (e.g., L2RPN) show that agents achieve reliable scores, and are resilient to adversaries; driving interest in real-world deployment. This paper argues that current evaluation settings understate real-world adversaries and introduces PEGAS, a query-based, gray-box attack framework that outperforms L2PRN adversarial evaluations. PEGAS scores candidate attacks using a Physics Informed Graph Attention Network surrogate for Optimal Power Flow solvers. PEGAS searches the mixed discrete–continuous, non-convex space with an elitist genetic algorithm, and projects candidates back onto feasible, connectivity-preserving actions before deploying them against static solvers and live RL-based grid-control agents. Our convergence analysis shows that PEGAS' best-found fitness converges almost surely to the global optimum. Empirically, on 118-bus grids and standard L2RPN benchmarks (IDF 2023, WCCI 2022), PEGAS substantially reduces defender survival rates compared to existing opponents and beats the winning approaches of these challenges. Specifically, PEGAS achieves a 93.2% success rate against static solvers and decreases the survival of the last L2RPN winner agent from 72% to 5.3%.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Reinforcement learning, Adversarial Learning, Robustness, Agentic Safety

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_feCVOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a new framework for finding high-impact contingency scenarios of a power grid using a genetic algorithm that is coupled with a surrogate model that approximates a physics based solver of the grid. The surrogate scores each candidate attack so the search avoids calling the solver every time. The attack space covers both line cuts and bounded load changes under a shared budget. The paper demonstrates advantages in two settings. Against static solvers the method identifies more contingency violations than the comparison methods. Against the dynamic L2RPN agents it replaces the challenge’s own opponent and drops defender survival well below what that opponent achieves, including against the 2023 challenge winner.

Critical strengths

This paper tackles a challenging problem in power grid operations of finding contingencies that threaten the stability of the grid. This matters because prior opponents in the L2RPN setup do not actively search for a defender’s weak points, so the agents were only tested against weak adversaries, and this work shows that gap clearly. The search design is reasonable for the combinatorial problem. The genetic algorithm handles a mixed discrete–continuous action space, line cuts and load changes together under one budget, and using a learned model as the fitness lets the search avoid a solver call per candidate. They also test in two settings and report an exhaustive upper bound so the gap to the ceiling is visible.

Critical weaknesses

• The RQ1 success metric is too weak. The paper measures success as the attacked state has more violations than the clean state, however there are instances in the experiments where one bus is barely over its limit but still counts the same as a collapse. This metric does not quantify severity of the contingency as a result, does not predict real impact of these outages. • The post-contingency state in RQ1 is set up in an unrealistic way. All non-slack generators are frozen at the surrogate’s predicted dispatch and the whole imbalance is put on the slack bus. Real post-contingency flow spreads the pickup across units through droop and participation factors. • RQ1 uses one set of limits with no emergency band. Voltage and thermal violations are checked against a single static bound. In operations there are separate normal and emergency limits, and overload is time dependent. A voltage excursion that is fine under emergency criteria gets counted here as a successful attack.

• Once concern is the accuracy of the surrogate itself. The search is guided entirely by the surrogate’s predicted violations, so if the surrogate is wrong, the attacks it selects are chosen on a flawed signal. This is reported in the manuscript as generator RPA sits around 0.74 and the authors admit it cannot respect tight reactive limits. In addition, the fitness has no thermal term at all even though line overloads are the real failure mode in RQ2. The accuracy also drops on the larger grid. This raises a deeper question about the evaluation. If the surrogate carries error, how can the success of PEGAS be judged against it, when the same model both selects the attack and underlies how impact is assessed? The paper never bounds the surrogate’s error or shows that it is small enough for the reported success rates to be trusted.

• There is no comparison to sensitivity based contingency screening (LODF/PTDF ranking), which is what operators actually use to find worst case outages without a solve per candidate. Since the results are driven by bridge cuts, a one shot LODF ranking might find the same lines much faster than a 500 generation GA. • Due to the way success is defined, the severity of the contingencies that were missed are never considered. The 6.8% in RQ1 and the 5 to 6% in RQ2 that survive are not analyzed. There is nothing on whether the missed cases are the critical ones, or whether the failed searches run longer because of the early stopping. • The improvement in runtime is only shown at 118 buses. The 6.34 s per attack is small, but it holds partly because the load genes (63 s on their own) are barely used at k=1. The 2000 bus case has no search timing and no dynamic run, so the “near real time, grid pace” claim is not backed at a realistic grid size.

Questions for authors

• In D.6, 613 of 727 collapses end on a line 26 disconnection, and line 26 is a bridge in the grid while RQ2 turns off the connectivity check. This suggests most of the result comes down to finding the one line whose removal islands the grid. Why is the full surrogate-plus-genetic-algorithm method needed to find this, when a simple graph check for bridges or cut vertices would identify line 26 directly at a fraction of the cost? Can you show how PEGAS performs against such a baseline, and how much of the reported success remains once the bridge cuts are set aside?

• Can you quantify the surrogate’s error on the quantity that actually drives the attack, i.e. how often its predicted violations match the power-flow ground truth? Given that generator RPA is ~0.74 and there is no thermal term, how do you know the search is being guided toward truly damaging actions rather than surrogate artifacts? • Since the surrogate both selects the attack and (in RQ1) underlies how the state is evaluated, how do you rule out that part of the measured success is the surrogate’s own error rather than real grid impact?

• Your success criterion only checks whether the number of violations increased with no indicator of severity. How many of your reported RQ1 successes correspond to large violation and can you report overload magnitude, voltage deviation size, or load shed?

7 comments
AuthorsOpenReview ↗

Severity and impact of the outages

We thank the reviewer for highlighting the importance of the problem and the significance of the gap we are addressing. We address below the weaknesses and questions you raised (grouped when overlapping).

R3W1/Q3: The success rate metric in RQ1 does not quantify severity of the contingency as a result, does not predict real impact of these outages.

Thank you for raising this discussion. Indeed, we are not distinguising between contingencies by severity in RQ1. We are using the common threshold in power systems literature: 1.05 PU.

The threshold we consider to rate the attack as successful is used by multiple official power grid referentials and standards for Transmission Grids Operations as the limit for safe operations: ENTSO-E parameters document[1], ERCOT voltage limits maintenance[2], SIMONA model documentation[3].

While we agree that lines can survive a few steps over 1.05 without grid collapse, the phenomenon is stochastic, and we are strictly following the safety threshold in RQ1. In RQ2, with the agentic simulator, the success criteria is the collapse of the grid.

  1. ENTSO-E. Parameters related to voltage issues (NC RfG implementation guidance document). https://www.entsoe.eu/Documents/Network%20codes%20documents/NC%20RfG/161116_IGD_Parameters%20related%20to%20voltage%20issues_for%20publication.pdf?Web=1
  2. VDE/FNN. Dielectric strength study (2020). Operation at 1.05·Um (relative to the highest equipment voltage Um, not nominal voltage). https://www.vde.com/resource/blob/2490180/3f0bf0d9d79ddbb01abe78b5fd1850fe/fnn-study---dielectric-strength--2020--data.pdf
  3. ERCOT. Voltage Limits Maintenance. Planning default limits: 0.95–1.05 (base case), 0.9–1.05 (post-contingency). https://www.ercot.com/files/docs/2014/03/14/voltage_limits_maintenance.pdf
  4. SIMONA documentation. Voltage Limits (cites EN 50160:2020-11). 380 kV EHV default: 0.9–1.05 p.u. https://simona.readthedocs.io/en/stable/models/voltage_limits.html
AuthorsOpenReview ↗

Realism of our study

R3W2: The post-contingency state in RQ1 is set up in an unrealistic way. Droop and participation factors are not considered.

RQ1 does not consider any corrective measure to the grid, or grids with "defense" mechanisms. We are following a static setting of "single slack", contrary to the "distributed slack" setting that allows participation factors.

Droop and participation factors are dynamic mechanisms that require considering primary frequency responses, which is out of scope for RQ1.

It is worth noting that the Grid2Op environment of the challenge in RQ2 solves a steady-state AC power flow at each 5-minute step. There is no frequency state and no governor dynamics, so droop is not modeled. RQ1 uses the same model and the same hypotheses as the challenge.

We will clarify the threat model of RQ1 and clearly scope our study to the "static single slack" model and explain the motivation of our setting (to match the target environment of RQ2).

R3W3: In operations there are separate normal and emergency limits, and overload is time dependent. A voltage excursion that is fine under emergency criteria gets counted here as a successful attack.

In RQ1, we are using only one threshold following well established power grid standards and referentials (see R3W1). Considering two thresholds is only relevant in multi-step time-dependent settings (as you state in the weakness), which is the setting of RQ2. RQ1 only considers independent single step snapshots.

In RQ2, the RL environment uses a multi-step simulator that considers this time dependency. The grid collapses if a line exceeds emergency limits immediately, or normal limits (limits used in RQ1) for multiple consecutive steps (stochastically).

Therefore, our single-threshold limit is perfectly suitable for the setting of RQ1. We will further explain this design choice in the threat model.

AuthorsOpenReview ↗

Evaluation & baselines

R3W4: One concern is the accuracy of the surrogate itself. If the surrogate carries error, how can the success of PEGAS be judged against it, when the same model both selects the attack and underlies how impact is assessed?

We would like to clarify that the success rate is never evaluated using the surrogate. An independent solver scores the success metric. In RQ1, we use a power flow solver implemented within PowerModels library, and in RQ2, the Grid2Op environment has its own solver to check the survival or collapse of the grid.

R3W5: There is no comparison to sensitivity based contingency screening (LODF/PTDF ranking)

Thank you for suggesting this non-AI baseline. Indeed, we only compared ourselves in RQ1 to Machine Learning based attacks. We have implemented this baseline for 118-bus and 2000-bus grids and our results show that within our threat model (same knowledge as PEGAS), this baseline yields 0% success rate both in the N scenario (in distribution) and the N-1 scenario (out of distribution):

GridMethodSuccess N (%)Success N-1 (%)
case 118PEGAS93.2 ± 3.086.3 ± 5.4
case 118LODF0.0 ± 0.00.0 ± 0.0
case 2000PEGAS71.6 ± 5.966.8 ± 6.1
case 2000LODF0.0 ± 0.00.0 ± 0.0

We will include these new baselines in the appendix.

Reviewer_fUYrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a framework, named PEGAS (Physics-informed Evolutionary Graph Adversarial Search), that generates strong, query-based adversarial attacks against Reinforcement Learning (RL) agents managing power grids.

To this end, the proposed PEGAS integrates a physics-informed Graph Neural Network (PIGNN) surrogate with a genetic algorithm. The PIGNN model approximates a full AC Optimal Power Flow (AC-OPF) solver while the genetic algorithm searches the mixed discrete-continuous attack space.

In the experimental analyses, PEGAS improves success and survival rates compared to the state-of-the-art in various benchmarks.

Critical strengths

First, the paper shows that current RL agents are fragile under strategic attacks. This is a nice observation which can guide designing safety-critical systems for safe deployment of RL.

Second, the paper shows that adding a physics-informed penalty on GNN surrogates is highly effective, while these pure data driven surrogates struggle when adversarial actions alter the grid topology. The experimental results show that this approach improves out-of-distribution generalization.

Critical weaknesses

The individual technical components of PEGAS have limited novelty. First, PIGNN surrogates for power grids are already established in the literature, and the proposed optimization method is a direct extension of existing evolutionary search algorithms.

Therefore, the primary contribution is the integration of these two existing tools within the Grid2Op simulator. This makes it more of an applied systems evaluation paper than a fundamental algorithmic breakthrough.

In the theoretical analyses, the paper presents theoretical convergence guarantees for the search algorithm. However, Theorem 1 only proves almost-sure convergence for the discrete, topology-only search space, while the discrete-continuous search space is addressed in Theorem 2. This theorem relies on a strict regularity assumption on the mutation kernel which remains unverified mathematically for realistic power flow constraints. Therefore, the theoretical support for the proposed hybrid algorithm is limited.

In addition, the authors report that baseline attacks like PGD, PRBCD, and GRBCD achieved a 0% attack success rate in their original forms. Then, an Action Validity Checker and Repair module was added to make these baselines functional. Therefore, the baseline performance depends on the quality of the repair module.

Questions for authors

As the size of the grid scales, the search space EvA grows exponentially.

How does the computational cost of the genetic search scale with grid size?

Does the evolutionary search become too slow or lose effectiveness on larger systems, even with the GNN surrogate?

3 comments

Thank you for your insightful feedback, and for recognizing the important insights and observation of our work and the effectiveness of our Physics-informed Graph Attention Network surrogate. We address the weaknesses you raised (Wxx) and your questions (Qxx) in the following:

R2W1: Limited novelty, and contribution is not strong enough to stand as a novel solution.

While our work does not introduce totally novel mechanisms, our contribution and novelty lies in the new research problem formulation we are introducing and the creative solution implemented to solve it:

1- We are first to define a realistic threat model for transfer attacks against power grids agents, and demonstrate that traditional adversarial attacks all fail with 0% success rate and require adjustments (novel insight, challenging common conceptions).

2- Next we demonstrate that using a Physics informed surrogate mitigates the incomplete information an attacker has on its target grid and demonstrate the effectiveness of Physics informed GNN against static power grids (novel mechanism for transferable graph attacks).

3- We set up a full agentic setting and demonstrate (1) that our evolutionnary search guided with Physics informed GNN can break the best agents (novel insight, challenging the way researchers are building RL agents) and (2) that the L2RPN benchmark is using very weak opponents, that are outdated and PEGAS is a strong alternative (novel benchmark to drive future evaluations & competitions).

Therefore, while the individual components exist in adjacent litterature, the combined mechanism, the insights, and the new research directions opened by our work are novel and significant.

AuthorsOpenReview ↗

Theory & assumptions

R2W2: The theoretical support for the proposed hybrid algorithm is limited.

In the main paper, we provided the theorem for the discrete space algorithm, and in the appendix the theorem for the hybrid algorithm:

  1. We only need the discrete algorithm for the agentic scenario (RQ2). Indeed the RL opponent can only use discrete actions. The assumptions, theorems and proofs are sound for RQ2 and empirically validated.

  2. The static solver scenario (RQ1) is evaluated using both algorithms in Figure 4: Using the hybrid algorithm PEGAS achieves 93.2% success rate. Restricting PEGAS to the discrete algorithm still achieves 91.6% success rate, ahead of the strongest baselines Eva (85.2%). This confirms that the assumptions, theorems and proofs are also sound for RQ1 and empirically validated.

  3. In the hybrid space of RQ1, the strict regularity assumptions needed for the proofs of the hybrid algorithm stand:

(1) Power-flow constraints never enter the search domain. The hybrid space is a finite set of edge and load choices times boxes of load deltas, B(z,I)=∏i∈I[−ϵ,ϵ]d∩[ℓi−xi, ui−xi]B(z,I)=\prod_{i\in I}[-\epsilon,\epsilon]^d\cap[\ell_i-x_i,\,u_i-x_i]: a product of intervals. AC power flow and the voltage and thermal limits enter only through the objective Φ\Phi. Its continuity (Assumption 1) holds by construction: a continuous neural network composed with hinge functions. The topology rules (i)–(iii) are a single repair applied after the search, outside the theorem.

(2) Assumption 2 is a property of our mutation operator, and it is proven. B.4.2 shows that EvA+'s uniform-then-clamp sampler satisfies it with c0=(2ϵ)−dc_0=(2\epsilon)^{-d}: clamping moves mass from outside B(z,I)B(z,I) onto its boundary and never removes mass from inside it. The assumption does not depend on the grid. In the revision we will state this as a lemma, so that Theorem 2 rests on a proven property rather than an assumption.

(3) We extend the guarantee from a fixed split to the full hybrid space. Crossover keeps the elite's split, and at k=1k=1 the split-shift mutation moves the elite into the other split in one step with positive probability. Hence EvA+ with hybrid actions (Algorithm 2) converges almost surely to the global optimum of the whole hybrid space.

Thank you for raising this critical point, we will clarify which theorem is used in each research question, and further explain why the assumptions hold in our setting through the new lemma.

AuthorsOpenReview ↗

Contribution of the repair mechanism

R2W3: The baseline performance depends on the quality of the repair module.

The authors report that baseline attacks like PGD, PRBCD, and GRBCD achieved a 0% attack success rate in their original forms. Then, an Action Validity Checker and Repair module was added to make these baselines functional. Therefore, the baseline performance depends on the quality of the repair module.

Correct; however, we would like to clarify that the repair module used in the baselines is the same as the one used in PEGAS. This demonstrates that PEGAS' physics informed graph attacks outperforms the baseline attacks (93.2% vs 16.0% for best baseline with repair) and is key to achieve this performance; the repair module alone is far from sufficient.

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

This paper proposed PEGAS, a query-based gray-box adversarial attack framework designed to assess the robustness of RL-based agents used in neural solvers for power grids. PEGAS considered benchmark and challenges as the baseline defense methods and devised the attack that can fool this agents. PEGAS targets the grid topology and load by utilizing the elitist genetic algorithm. It considers physics-informed Graph Attention Network as the surrogate model for the power grid OPF solver. Evaluation results show that PEGAS achieves high attack success rate on a 118-bus system against both static and agentic solvers.

Critical strengths

Designing the genetic algorithm to search over the complex search space to find the solution that achieves the attack's objective. Besides, consider the physics-informed Graph Attention Network as the grid surrogate model provides a unique strength of consider a dynamic power grid non-linear in an efficient way.

Critical weaknesses
  • Threat model is not clearly defined or in another word contradictory. One of the core challenge to solve is to design an adversary who can attack a power grid with incomplete knowledge (C2) still it is assumed that the attacker has trained a model with a representative dataset that incorporate the complete behavior of the systems.
  • In another challenge (C1) the attacker is supposed to be operating in near-real-time. But the proposed attack takes (6.34–57.8s) to generate the attack which is not practical for power system SCADA where the data is transmitted more frequently.
Questions for authors
  • The evaluation heavily centers on competitive benchmarking challenges (e.g., L2RPN). Could the authors reframe the motivation and attack contribution from a practical utility operator's perspective, specifically who does not participate in these benchmarks, and why they should be concerned about PEGAS. Besides, how does this framework fundamentally advance the robustness analysis of real-world AI-enabled power systems beyond existing stress-testing tools?
  • The threat model (C2) assumes an adversary with incomplete system knowledge. However, PEGAS relies on a pre-trained physics-informed Graph Attention Network surrogate that must accurately capture system behavior. Can the authors clarify how an attacker acquires a sufficiently representative dataset to train this surrogate model under realistic, incomplete-knowledge constraints without triggering system alarms?
  • The paper highlights a core challenge (C1) of designing a near-real-time adversary. However, the evaluation notes that generating an attack takes between 6.34 to 57.8 seconds for a relatively small 118-bus system. Given that real-world power grids contain thousands of buses and require sub-second or rapid response cycles, how does PEGAS scale to larger topologies?
5 comments
AuthorsOpenReview ↗

Clarification of the threat model

We thank the reviewer for the constructive feedback, and the praise about the design of our attack and search algorithm. We address the weaknesses you raised (Wxx) and your questions (Qxx) in the following:

R1W1 + R1Q2: Incomplete or contradictory threat model (access to information / surrogate design)

We believe there is some misunderstanding. Our threat model assumes the attacker has NO access to neither the target grid nor its dataset. The surrogate is trained on grids provided by OPFData that are significantly different from the target grids (please see table below).

The design of this realistic threat model is actually one of our novelties compared to other (less realistic) adversaries. Indeed, the baselines opponents collect the observation given from Grid2Op reinforcement learning framework, which represents the complete state of all the nodes of the target grid.

Comparison between target and surrogate grids:

AspectGrid2Op (Target Grid)OPFData (Surrogate grid)Effect on the attack
Buses / substations118118Matches.
Loads9999One-to-one by bus; load mapping is exact.
Generators62 production units54 generator nodesThe surrogate is not aware of 8 target generators.
Storage7 units, separate from generatorsnoneThe surrogate is not aware of the batteries and their charging state.
Lines / transformers186 line ids, 179 unique endpoint pairs175 line entries (168 unique endpoint pairs)The surrogate is not aware of a subset of lines.
Branch featuresGrid2Op thermal limits, 5-min dynamicsstandard PGLIB r, x, b line attributesWe assume the target grid uses standard "power lines" properties
Bridges915Bridges are critical branches whose removal increases the number of connected components.

As the table shows, only an approximate knowledge of the grid is available to the attacker. We will clarify section 3.2 and add this table to Appendix A.

AuthorsOpenReview ↗

Motivation and scalability

R1Q1: Motivation - Reframing the problem from a practical utility of PEGAS from the operator's perpspective

PEGAS is built against a benchmark designed by a power grid operator, RTE, the national power grid operator in France, in collaboration with the dutch power grid operator. It aims to (1) develop next generation of decision-making methods powered by supervised and reinforcement learning for critical infrastructures, and (2) ensure trustworthiness by design of these solutions.

The tooling, data, and simulation engine were designed to mimic Paris region power grid and represent real operation scenarios. PEGAS brings two benefits: 1) it allows to stress test RL operation agents in extreme and OoD scenarios and 2) Evaluate the impact of physical grid attacks (spoofing, line cutting, load boosting) with a fast testbed. The adversarial scenario we are bulding was considered in the NeurIPS 2020 challenge of the competition, and was since never improved. Meanwhile power grid agents have beaten these baselines and are now being considered by multiple power grid startups and commercial software (Kyra, DigSilent Power Grid, NEPLAN 360) all are integrating AI based grid operators.

R1W2 + R1Q3 : The proposed attack takes (6.34–57.8s) to generate the attack which is not practical for power system SCADA where the data is transmitted more frequently.

Power grid operation refer to near-real time as actions taking less than one simulator step to run, i.e., less than 5min. This is currently the case for our attacker, and is the setting of the L2RPN challenges we are targetting.

PEGAS meets the definition of near-realtime with a median of 1.9 s on IDF 118 buses. The cost is dominated by surrogate forward passes, which scale about linearly with edges.

We cannot evaluate PEGAS speed on a much larger power grid against real agents, because (1) there is no Power Grid RL environment in Grid2Op for grids larger than our current 118 bus grid, (2) no public RL agent was trained on Grid2Op for larger grids and it would be unfair to attack agents designed for 118 bus on much larger grids.

All other RL for power grid benchmarks (Gym-ANM, MARL2Grid) have similar or smaller grid size. PEGAS was evaluated on the SoTA benchmark of the field. 118 buses represents already a challenging agentic setting with more than 65,000 discrete actions to handle.

We ran the Physics Informed GNN surrogate and the PEGAS against static solvers (RQ1), that do not require an RL environment, and our results in the following table show that both scale linearly with the size of the grid. Even for a 10k grid, PEGAS runs within the 5min limit.

Mean timeCase 118Case 2000Case 10000
GNN surrogate0.071 ms1.01 ms4.95 ms
PEGAS3.68s16.9s60.8s

We will expand the discussion on scalability and higlight that PEGAS is limited by the benchmarking environments and not by any of its properties.

Reviewer_aeKKOpenReview ↗

Assessment of Authors’ Rebuttal: Key Concerns and Remaining Limitations

Thank you for the detailed rebuttal. My assessment of the key concerns is as follows:

  1. Threat Model (Partially Addressed). The clarification on surrogate training is helpful. However, the exact bus and load correspondence between the surrogate and target grids raises questions about realistic attacker knowledge.
  2. Practical Relevance (Largely Addressed). The connection to RTE's operational challenges strengthens the motivation, although applicability beyond benchmark environments remains unclear.
  3. Near-Real-Time Feasibility (Partially Addressed). The 5-minute operational window provides reasonable justification. However, the discrepancy between the originally reported runtimes (6.34–57.8 s) and the new 1.9 s median needs clarification.
  4. Scalability (Largely Addressed). The additional 2,000- and 10,000-bus experiments demonstrate computational scalability, but only against static solvers, not RL-based agents.

Overall Assessment: The rebuttal addresses several concerns, particularly scalability and motivation. However, questions about realistic attacker knowledge and practical applicability remain.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Summary

We thank the reviewers for their questions and feedback. The revised PDF includes:

Main text

  • Sec. 3.2: clarified the attacker's knowledge (no access to the target grid's model or dataset), with a pointer to the new grid comparison (A.4)
  • Sec. 4.4: the convergence guarantee now also covers the whole hybrid search space at the budget k=1k=1 used in our experiments (Theorem 3)
  • Sec. 7: pointer to the new scalability results, up to a 10,000-bus grid (D.13)

Appendix

  • A.3: relevance of the L2RPN benchmark for grid operators
  • A.4: target grid vs. the surrogate's training grid (what the attacker knows)
  • B.4.2: Assumption 2 is now a proven property of the EvA+ sampler (Lemma 3); convergence over the whole hybrid space at k=1k=1 (Theorem 3); which result covers each research question (Remark 2)
  • C.1: limits and slack model of the static setting (single slack, PGLIB ratings, time-dependent limits handled in RQ2)
  • C.4: protocols of all new studies (C.4.1–C.4.6)
  • D.2: the baselines use the same repair as PEGAS (16.0% vs. 93.2%)
  • D.6: comparison with sensitivity-based contingency screening (LODF) on case118 and case2000 (0% for LODF)
  • D.9: bridge cuts: success of PEGAS without bridge cuts (98.4% in RQ1; 44.1% mean defender survival in RQ2)
  • D.10: analysis of the missed cases: why PEGAS misses them, whether they are the critical ones, and whether the search stops too early
  • D.11: the surrogate's error and attack success
  • D.12: severity of the successful attacks: success at higher thermal limits (up to 2×), violation sizes, corrective re-dispatch, and energy not served in RQ2
  • D.13: search time on case118 and case2000, and scaling up to a 10,000-bus grid

These additions confirm the main results of the paper, which remain unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.