acceptodds
Under review as a conference paper at ICLR 2027

LLM-Assisted Clean-Label Data Poisoning of Language Models

Abstract

Fine-tuning on externally sourced data exposes language models to data poisoning: an attacker manipulates training examples to influence the learned model. Clean-label backdoor attacks preserve the original labels while inducing targeted predictions on inputs carrying an attacker-chosen trigger. We introduce LLM-CT, an LLM-assisted method that automates dataset-specific candidate generation and combines sentence screening with low-confidence host selection. Shared prompts guide an LLM to infer corpus-specific generation fields and carrier topics, producing reusable sentences for lexical and clean-reference behavioral screening. An accepted sentence is appended to selected target-class hosts, preserving their text apart from normalization. Construction requires no victim queries, weights, or gradients. We report attack success rate (ASR) alongside ASRD, its signed difference from a matched clean-trained control, to distinguish poisoning-induced effects from pre-existing target responses. In the rate study, LLM-CT achieves 95.9–100% ASR at 1% poisoning across BERT, GPT-2, and Llama-2-7B on SST-2, OLID, and AG News. Corresponding ASRD values of 92.9–98.6 percentage points demonstrate substantial poisoning-induced effects. Several baselines instead exhibit near-zero ASRD in some settings, indicating that their raw success largely reflects responses already present before poisoning. Fixed-trigger controls further demonstrate the value of low-confidence host selection at small budgets. Under seven individually applied BERT defenses, evaluated constructions retain 92.76–100% ASR at poisoning rates of 0.5–1%. Defense results show that partial poison removal and input rewriting can leave attack success high, motivating evaluation of post-defense behavior alongside detection rates and clean-data costs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Data poisoning, backdoor attack

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_iBukOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces LLM-CT, a clean label poisoning method for text classifiers. It uses an LLM to generate reusable sentence triggers, screens them for lexical constraints and their effects on a clean reference model, and appends an accepted sentence to target-class examples with low confidence under an ELECTRA surrogate. The experiments cover SST-2, OLID, and AG News with BERT, GPT-2, and Llama-2-7B victims. The paper reports high attack success at small poisoning budgets, compares against matched clean-trained controls using signed ASRD, and evaluates seven defenses on BERT. Additional experiments examine host selection and the candidate acceptance threshold.

Critical strengths
  1. The reported attacks are effective at small poisoning budgets across several tasks and model architectures. At 1% poisoning, LLM-CT reaches 95.9–100% ASR across all nine dataset and victim combinations, while clean accuracy stays within 1.1 percentage points of the matched clean control. These results provide useful evidence that a small number of examples containing an ordinary sentence can create a strong backdoor, including when the victim differs from the models used to construct the attack.

  2. The evaluation goes beyond reporting raw ASR. The matched clean controls distinguish effects learned during poisoning from target responses already caused by the trigger. The fixed-trigger host-selection experiment also provides a useful controlled comparison. In the defense analysis, reporting intervention counts alongside post-defense ASR helps explain why identifying or modifying some inputs does not necessarily remove the attack.

Critical weaknesses
  1. The experiments do not sufficiently isolate the contribution of LLM-assisted sentence construction. Sentence insertion is established, and the Minimum host-selection strategy is explicitly adopted from Kirci and Gursoy (2026). Figure 2 shows that host selection matters for one generated sentence, but it does not show how much the proposed generation procedure adds when host selection is held fixed. The AddSent comparison changes several factors at once: Table 7 uses a fixed movie-related sentence inserted at a random position, whereas LLM-CT uses corpus-adapted suffixes and low-confidence hosts. A comparison using conventional sentence triggers with the same hosts, insertion position, and budget would make the contribution much clearer. Without this control, it is difficult to determine how much of the improvement requires the proposed construction procedure.

  2. The evidence for the automated pipeline is split across different configurations and evaluation protocols. Appendix A.6 states that the inferred-prompt configuration uses benchmark test inputs for reference screening and has not been evaluated on decoder victims or under defenses. Its accepted sentences in Table 13 also differ from the rate-study sentences in Table 6. Therefore, the strongest transfer and defense results do not directly validate that particular automated configuration, and its reported outcomes are development evaluations rather than independent test results. Appendix A.5 provides a useful experiment with screening data separated from the test set, but only for the SST-2/BERT threshold ablation. The main rate and host-selection comparisons also use one training seed. These limitations leave uncertainty about whether the full automated procedure reliably reproduces the headline results.

I lean toward rejection because the current evidence does not yet establish the added value and reliability of the automated construction procedure. The strong reported attack results make these questions worth resolving.

Questions for authors
  1. Which generation prompts, screening data, and selection rules produced the sentences used in the main rate and defense studies? How do these configurations differ from the inferred-prompt procedure in Appendix A.6? Please clarify whether any test inputs influenced construction or configuration selection for each main experiment. A small repeated evaluation of the complete automated procedure with test data excluded from construction would help address this concern.

  2. How does LLM-CT compare with conventional sentence triggers when both use the same ELECTRA-ranked hosts, suffix position, and poisoning budget? A focused comparison at the low poisoning rates where the claimed advantage is largest would help distinguish the contribution of sentence construction from that of host selection.

  3. Is there an independent assessment of whether the appended sentences preserve the original semantic labels? Appendix A.6 correctly notes that low target-directed reference leakage does not establish semantic label preservation. A manual check of a sample of the selected low-confidence hosts and their modified versions would help substantiate the clean label claim.

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

The paper proposes LLM-CT, a clean-label text poisoning method that generates and screens reusable sentence triggers and appends them to low-confidence target-class training samples. It evaluates attack effectiveness across three datasets and three victim models, using matched clean-trained controls and several backdoor defenses.

Critical strengths
  1. The paper addresses a relevant training-data security problem.
  2. The evaluation covers multiple datasets and models, and matched clean-trained controls help interpret the reported attack effects.
Critical weaknesses
  1. The incremental benefits of the proposed generation and screening pipeline are insufficiently established. The experiments demonstrate feasibility and overall attack effectiveness, but do not adequately isolate the contributions of the new components or establish their advantages over simpler trigger construction methods.
  2. The comparison with recent, directly relevant clean-label attacks and defenses is incomplete. This limits the evidence for the method’s competitiveness and the scope of its defense-resistance conclusions.
Questions for authors
  1. Please provide controlled comparisons with simple neutral-sentence triggers and ablations of the generation and screening components, particularly LLM review and lexical filtering. Keep hosts, poisoning budgets, and trigger placement fixed, and report ASR, ASRD, label consistency, and post-defense performance to establish the incremental benefits. There is currently a lack of evaluation for the sentences that have been filtered out. Therefore, you did not verify your contribution.
  2. Lack of the latest baseline. Please consider updating the attack baselines to include AttrBkd [1], EST-Bad [2], and CleanBattack [3], with appropriate matching of threat models and poisoning budgets or other possible baselines.
  3. Please consider BMC [4] as a training-time defense and X-GRAAD [5] as an inference-time defense or other possible baselines.
  4. Please provide the complete forbidden-stem construction procedure, including tokenization and stemming, the Wilson confidence level, and the class-association threshold. Please also report rejection counts for each lexical rule to clarify which constraints dominate candidate filtering.

[1] Wencong You and Daniel Lowd. The Ultimate Cookbook for Invisible Poison: Crafting Subtle Clean-Label Text Backdoors with Style Attributes. SaTML, 2025.

[2] Ziqiang Li et al. Large Language Models Are Good Attackers: Efficient and Stealthy Textual Backdoor Attacks. Pattern Recognition, 2026.

[3] Huahui Li et al. CleanBattack: A Clean-Label Text Backdoor Attack with Limited Information. PMLR 278, 2025.

[4] Zhenting Wang et al. Data-centric NLP Backdoor Defense from the Lens of Memorization. Findings of NAACL, 2025.

[5] Anindya Sundar Das et al. Unmasking Backdoors: An Explainable Defense via Gradient-Attention Anomaly Scoring for Pre-trained Language Models. ICLR, 2026.

Reviewer_fy1kOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

They introduce an approach to LLM based data poisoning that identifies a single reusable trigger sentence that can be appended to selected training examples while leaving their labels unchanged. Repeated exposure reinforces the association between the trigger and the attacker’s target label. The method uses an LLM to generate dataset specific candidates, filters them to reduce target bias, and inserts the selected trigger into low-confidence training examples to increase its effectiveness. The authors evaluate whether these triggers transfer across model architectures and remain effective against several existing defenses.

Critical strengths
  • Broad evaluation across datasets, architectures, and baselines.
  • Matched clean-model controls distinguish genuine poisoning effects from pre-existing trigger bias.
  • Useful ablations and detailed implementation information support the main findings.
Critical weaknesses
  • Limited novelty over existing LLM-assisted data-poisoning methods (e.g., You et al. (2023))
  • Unclear problem statement and motivation.
  • Evaluation relies on relatively outdated target models used as conventional classifiers.
  • A single global trigger creates an obvious repeated-text signature. The paper does not evaluate whether a simple duplicate-sentence or n-gram-frequency defense could detect it.
Questions for authors
  1. Why do you evaluate only older models used as classifiers (BERT, GPT2 ...) instead of one of many newer architectures, and do you expect the attack to generalize to modern instruction-tuned LLMs?
  2. Did the authors at least discuss how this approach could be more easily defended against because it involves insertion of a single repeated attack phrase?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.