acceptodds
Under review as a conference paper at ICLR 2027

Backdoor Robustness in Language Models Across Trigger and Behavior Characteristics

Abstract

Backdoor attacks pose a significant security risk for large language models (LLMs), yet existing detection and mitigation methods are typically evaluated in simplified settings. We have 5 specific trigger instantiations with 5 behaviors. We do not isolate abstract trigger and behavior characteristics. I'd suggest: We introduce a benchmark covering 25 combinations of five trigger instantiations and five target behaviors covering various lexical and semantic characteristics. We consider a resource-constrained attacker using limited attacker-controlled data and modest compute, and evaluate representation-based detection methods alongside training-time (DPO) and inference-time (CleanGen) defenses. Our results reveal four key findings. First, diverse backdoors can be effectively implanted with limited data and compute. Second, detection performance varies substantially across trigger–behavior configurations, with existing detectors exhibiting particularly poor performance for topic-based triggers. Third, DPO can reduce backdoor behavior in some configurations but is highly dependent on the trigger–behavior pair, whereas CleanGen provides more consistent mitigation across most settings. Finally, association analyses show that attack success is associated with several dataset-level properties, while defense and detection outcomes are substantially less predictable; mechanistic analyses further reveal that backdoors vary in how diffusely they are represented across model components, with localized backdoors being more amenable to targeted steering. Together, these findings expose important limitations of current backdoor defenses and motivate approaches that are less dependent on known backdoor signatures and more informed by the mechanisms underlying misaligned behavior.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
backdoors, llm, safety, misalignment, data poisoning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_vKVbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduce a two-dimensional benchmark that systematically varies lexical and semantic diversity of both trigger and behavior and evaluate representation-based detection methods, DPO alignment and one inference-time defenses. The association analyses show that attack success is associated with several dataset-level properties, while defense and detection outcomes are substantially less predictable; mechanistic analyses further reveal that backdoors vary in how diffusely they are represented across model components, with localized backdoors being more amenable to targeted steering.

Critical strengths

The experiment results offered some useful information. The analysis of how distributional differences between backdoored and clean examples relate to attack success, defense efficacy, and detectability. These associations provide useful context for interpreting performance differences across datasets and models.

Critical weaknesses

The contribution of this paper is limited and narrow in scope. The evaluation focuses on conventional backdoor configurations and a limited set of detection and defense methods, with defenses only covering general DPO and CleanGen. It remains unclear whether the findings generalize to more sophisticated attacks or a broader range of countermeasures.

The current experiment is not sufficient to support some of the claims made in the article. For example, the conclusion that “preference-based safety training is not a reliable backdoor defense” too broad given the current DPO setup. The results do not clearly distinguish whether the observed failure stems from the preference optimization objective, the coverage of the training data, the capacity of the LoRA updates, or the optimization settings. And some of the hyperparameter choices and experimental settings are not clearly explained or analyzed.

Questions for authors

see Critical Weaknesses above

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

This paper presents a study on the robustness of detection and mitigation methods of backdoors in two LLMs (Llama 3.1 8B and Qwen3-4B). The authors vary the backdoor behavior and trigger for a total of 25 combination with the trigger spanning: prefix, sentiment, language, dialect and topic; and the behavior spanning: fixed refusal, fixed ad insertion, contextual refusal, contextual ad insertion and editorial biais. All backdoor are injected using LoRA (r=8) training. They show that backdoors can be injected with limited ressources.

They explore 4 existing detection methods (diff of means, mahalanobis distance, topological evolution dynamics and beatrix) and 2 existing mitigation methods (DPO and CleanGen). They show that detectability vary a lot between methods and backdoors and that mitigation using CleanGen is more effective than DPO.

Additionally, they do a mechanistic study of the backdoor models showing that backdoors cluster in specific attention heads are more amenable to steering.

Critical strengths
  1. The paper study a wide range of backdoor (trigger/behavior pair) across multiple models.
  2. The paper evaluate a substantial number of mitigation and detection methods.
  3. The paper study mechanistically backdoor in LLMs.
Critical weaknesses
  1. The findings that backdoor can be injected in a LoRA setup or with few examples is redondant with previous work [1, 2].
  2. The paper is an evaluation of existing method and do not propose new mitigation or detection methods. The contribution of the paper are therefore mostly analytical. Although this type of reproducibility study is important and necessary for science, I do not think that this paper pass the bar for top-tier conference as ICLR.

Typo in abstract:

  • I’d suggest: We introduce a benchmark covering 25 combinations of five trigger instantiations and five target behaviors covering various lexical and semantic characteristics.

[1] Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples; Souly et al.; 2025

[2] The Philosopher's Stone: Trojaning Plugins of Large Language Models; Dong et al.; 2024

Questions for authors
  1. Do you think that your mechanistic interpretability findings hold in a standard SFT injection setup ?
  2. Do you think that results would change with a harmful behavior in a backdoor ?
Reviewer_kddQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper builds a backdoor benchmark for LLMs that varies two things at once: the trigger (a fixed prefix token, polite rewrites, translating the prompt into Turkish, Gen-Z slang, and a politics topic condition) and the target behavior (fixed vs. contextual refusals, fixed vs. contextual ad insertion, and right-leaning editorial bias), crossing them into a grid of configurations. Backdoors are implanted by LoRA fine-tuning with a small amount of attacker-controlled data and modest compute on Llama-3.1-8B and Qwen3-4B, with checkpoints chosen so that general capability stays close to the base model. The authors then test how this plays out for defenses: several activation-based detectors (Diff-of-Means, Mahalanobis, TED, Beatrix) and two mitigations (DPO on safety data, CleanGen at inference time), plus correlations between dataset statistics and outcomes, a causal mediation / steering study on a smaller Llama model, a model-scale study, and a human study validating the LLM judges. In short, backdoors turn out to be easy to plant but neither detection nor mitigation is reliable across the board: detectors mostly fail on topic triggers, DPO helps on some configurations and not others, CleanGen works almost everywhere except editorial bias, and backdoors that are localized to a few heads can be steered away.

Critical strengths
  1. Using low-resource LoRA fine-tuning, the paper evaluates four detectors and two mitigation methods across 25 trigger–behavior pairs, with additional experiments across model scales. The results provide useful evidence that performance on one configuration may not transfer to others.
Critical weaknesses
  1. Limited explanation of the empirical findings. Section 7 provides correlations rather than causal explanations, while Section 8 uses a different, smaller model and evaluates steering on only four configurations. These analyses therefore do not adequately explain the configuration-dependent detection and mitigation results across the 25 settings.

  2. Overstated novelty. Prior work already goes beyond fixed token triggers and deterministic responses, covering tasks such as jailbreaking, misclassification, and targeted refusal. For example, BackdoorLLM spans eight attack strategies, seven scenarios, and six model architectures, while CROW evaluates six poisoning attacks. The authors should clarify their incremental contribution over these benchmarks.

  3. Each trigger category has only one instantiation: Turkish for language, Gen-Z slang for dialect, politics for topic, and one polite rewrite for sentiment. Thus, the results concern five specific implementations rather than general trigger characteristics. Additional instantiations are needed to establish whether the difficulty ordering in Figure 2 generalizes.

  4. The topic-triggered configurations are confounded. Political articles serve as triggered inputs and other topics as clean inputs, so detector performance may reflect domain shift rather than backdoor activation. Some refusal settings also exhibit high ASR on clean, non-triggered inputs, indicating poor trigger specificity and making raw ASR misleading. The authors should provide appropriate controls and show whether Findings 2 and 3 hold without these configurations.

  5. No code or data is provided with the submission, only a promise of future release. For a benchmark paper, this limits verification and reuse. An anonymous repository containing the data and evaluation code should be provided.

Questions for authors
  1. Why is the clean-input attack success rate so high for the topic-triggered refusal configurations, nearly matching the backdoored models? Were clean political examples being counted as refusals by the verifier? If so, how do you tell backdoor activation apart from the base model's own refusal behavior in these rows, and would the main findings still hold if the topic rows were excluded?

  2. For topic triggers the detector only sees inputs that differ in article topic. What makes you confident the probe results there are about the backdoor rather than about topic shift?

  3. How sensitive is the trigger ordering to the specific instantiations? For example, would swapping Turkish for another language change much in Figure 2?

  4. The related-work section says you evaluate a broader set of mitigations than prior benchmarks, but the main text only has DPO and CleanGen. Did you try anything like BEEAR or latent adversarial training? I am curious whether the 'safety training is unreliable' conclusion holds against defenses that were actually designed with backdoors in mind.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewers

We thank the reviewers for their constructive feedback. We plan to improve our paper by including multiple instantiations of each category, add safety training baselines such as BEEEAR and strengthening our analysis.

We would also like to clarify that we provided the data and code as supplementary materials with the original submission. Our benchmark extends prior work by systematically varying both trigger and behavior characteristics and studying the resulting backdoors through detection, defense, and mechanistic analysis.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.