acceptodds
Under review as a conference paper at ICLR 2027

Two-Stage Data Poisoning Attack against the Large Language Model for Power Systems

Abstract

Large language models are increasingly being explored as decision-support components in power-system operation, but their dependence on data-driven learning introduces security risks during training and fine-tuning. This paper studies a stealthy data poisoning threat against a Transformer-based language-model-assisted power-system decision pipeline. We propose a two-stage state-aware attack framework. First, a local surrogate model and Jacobian-based augmentation are used to characterize sensitive input dimensions and local decision behavior. A multidimensional vulnerability index is then used to identify critical operating periods, while surrogate self-attention scores are used to prioritize critical nodes. Second, a low-rate poisoning strategy is formulated as a constrained bilevel optimization problem that jointly considers physical feasibility, statistical consistency, and attack effectiveness. An augmented-Lagrangian procedure is used to solve the resulting constrained problem, followed by sparse injections aligned with the system’s operating cycle. Experiments on an IEEE 39-bus simulation with a DeepSeek-R1- Distill-Qwen-1.5B-based decision pipeline evaluate target localization, detection evasion, and physical degradation. The reported results show that state-aware localization concentrates the observed degradation on the selected spatio-temporal targets, while the proposed constrained poisoning remains below the evaluated bad-data and distributional detection thresholds.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Data Poisoning Attack, Power Systems, Adversarial Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_VmLXOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a two-stage attack against a language-model-assisted power-system pipeline. It selects attack periods and nodes using a Transformer surrogate, a vulnerability index, and attention scores, then generates constrained perturbations. Experiments on an IEEE 39-bus simulation report two overload observations and 0% detection by the evaluated detectors.

Critical strengths

Table 1 separately varies attack nodes and operating periods. This is a useful initial control design, although the reported evidence is insufficient to draw strong conclusions.

Critical weaknesses
  1. The central attack and optimization method are not established. The paper does not specify which victim model learns from poisoned samples, its training objective, or its update schedule (§3.2 and §5.1). Therefore, the reported outcomes cannot be attributed to training-data poisoning. Equations (25)–(30) contain an objective and constraints but no lower-level optimization problem. Equations (31)–(33) also omit the variable splitting and alternating primal updates needed to specify ADMM. These omissions prevent verification and reproduction of the main method.
  2. The experimental evidence is inadequate. The attention-to-node mapping in Equation (24) is undefined. Table 3 reports detection rates without trial counts, repeated-run variability, or demonstrated budget matching across methods. Table 4 reports two overload observations on the same line, but no matched unattacked or frozen-model control establishes an accumulated learning effect. The claimed adaptation to changing topology is not evaluated.
Questions for authors
  1. Which victim component is updated using poisoned data, and what are its training objective and update schedule?
  2. What are the complete lower-level optimization problem and implemented ADMM updates?
  3. Can you provide matched attack budgets, repeated trials, and unattacked/frozen-model controls supporting the reported results?
Reviewer_GnrLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a two-stage poisoning attack against an LLM-assisted power-system pipeline. It first selects vulnerable periods and nodes, then generates perturbations intended to cause harmful dispatch decisions while evading detection. In simulation the authors report two overload events and 0% detection by the tested detectors.

Critical strengths

The paper addresses an important safety/security problem in LLM-assisted power systems, state-aware selection of attack locations and times is potentially interesting (although already explored in the power system security literature). The spatial and temporal control groups are also useful.

The problem formulation considers physical constraints and anomaly detection.

The paper compares a set of selected and alternative attack locations and periods.

Critical weaknesses

The motivation of the paper is understandable, but the end-to-end pipeline is unclear. The paper does not adequately connect forecasting, dispatch, the LLM, detection, and grid operation. Some technical terms, model inputs and outputs, and training procedures are insufficiently defined.

The attack mechanism is unclear. The paper describes training-time poisoning but does not explain when samples are injected, how the victim is retrained or fine-tuned, or which parameters are updated.

The model pipeline is not defined. It is unclear which model the surrogate approximates, what a clean and a poisoned sample contains, and how the LLM’s outputs become actions in the grid simulator. The relationship between poisoned learning data and the reported overloads is therefore hard to establish.

The considered optimization problem is not genuinely presented as bilevel. The equations describe one constrained objective, without a separate lower-level training problem.

The experiments lack essential details, training and optimization settings, uncertainty estimates. The paper motivates adaptation to changing topologies but it does not seem to perform an evaluation under changing topologies.

Questions for authors
  • What does one clean training sample contain, how is its poisoned version constructed, and when is it added to the victim’s training data?

  • Which victim model does the surrogate approximate, and how are model outputs converted into grid-control actions?

  • What is the explicit lower-level optimization problem considered?

  • What poisoning budgets, sample counts, repeated trials, and topology-changing experiments support the reported results

Reviewer_qo3GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a two-stage data-poisoning framework for an LLM-assisted power-system decision pipeline, combining vulnerable-condition identification with constrained, low-rate data manipulation. Experiments on the IEEE 39-bus system assess target localization, detection evasion, and operational degradation.

Critical strengths
  1. The paper addresses an important security concern: the vulnerability of LLM-assisted power-system decision pipelines to manipulated data.

  2. The framework considers both physical feasibility and statistical detectability, while its localization experiments examine the combined importance of attack location and timing.

Critical weaknesses
  1. The paper does not clearly establish whether the reported failures result from training-time poisoning or inference-time input corruption. It should explain how manipulated samples enter the victim model’s training or fine-tuning process and demonstrate that the resulting parameter updates cause degraded decisions on clean test inputs. Without this evidence, the central claim of a training-time poisoning attack remains insufficiently supported.

  2. The evaluation is limited to the IEEE 39-bus system, leaving the framework’s scalability and generalizability insufficiently established. Experiments on larger power networks such as IEEE-8500 bus, together with runtime, memory consumption, and performance across different operating conditions, would help assess its applicability beyond a small test system.

  3. The temporal evaluation scope is unclear. Although the paper describes vulnerability identification using weekly operation data, it does not report evaluation across multiple independent weeks or different temporal operating regimes, limiting evidence of temporal generalizability.

  4. The candidate-node set for target localization is insufficiently specified. Although the paper reports three selected buses, it does not clarify whether all 39 buses are considered or whether selection is restricted to a subset. This ambiguity limits assessment of the localization procedure and the realism of the attacker-access assumptions.

Questions for authors
  1. How do manipulated samples enter the victim model’s training or fine-tuning process? What evidence demonstrates that the resulting parameter updates degrade decisions on clean test inputs, rather than failures arising from inference-time input corruption?

  2. How does the framework scale to larger power networks in terms of runtime, memory consumption, and attack effectiveness? Does its performance remain consistent across different network configurations and operating conditions?

  3. Have the findings been validated across multiple independent weeks or different temporal operating regimes?

  4. Does target localization consider all 39 buses or only a restricted subset? What restrictions arise from measurement availability and attacker access, and how are these reflected in the candidate-node set?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.