acceptodds
Under review as a conference paper at ICLR 2027

Clean Data, Dirty Secrets: Backdoor Transfer in LLM Distillation via Defect Amplification

Abstract

Knowledge distillation (KD) is widely used to distill knowledge from large language models (LLMs) to compact student models, enabling efficient deployment in resource-constrained scenarios. It is commonly considered that using clean distillation data can mitigate the transfer of explicit trigger-based backdoors from poisoned teacher models to student models. In this paper, we propose DefAttack, a novel backdoor attack that enables backdoor transfer even when the student is distilled on clean data. Unlike existing explicit trigger-based backdoors that rely on transferable trigger-response associations, DefAttack exploits latent model defects that manifest as systematic biases in the teacher model's output logits. Specifically, we introduce a defect amplification loss during teacher poisoning, which increases the logits assigned to target-domain tokens while constraining changes to benign output logits. This transforms latent defects into transferable backdoors, allowing student models to implicitly inherit them while aligning with teacher supervision on clean distillation data. Importantly, DefAttack decouples backdoor transfer from activation: the defect backdoor is inherited during clean-data distillation and activated only afterwards using an optimized suffix. Evaluations across four models and three tasks demonstrate an average attack success rate of at least 95.3% and a false positive rate of at most 4.5%. Further analyses show that defect backdoors rely primarily on transferred defects rather than the optimized suffix.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Knowledge Distillation, Backdoor Attacks, Large Language Models, Model Security, Backdoor Transfer

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1R9GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DefAttack, aiming to transfer backdoors from compromised teachers to students through clean-data distillation. It amplifies target-domain token preferences in the teacher and subsequently activates the inherited preferences using suffixes optimized on surrogate students. Experiments cover four model families, three tasks, and multiple distillation strategies.

Critical strengths
  1. The paper investigates a meaningful model supply-chain risk that clean distillation inputs do not guarantee trustworthy teacher supervision.

  2. The experiments are thorough, covering four model families, three tasks, and multiple distillation strategies.

Critical weaknesses

Under the claimed soft-target transfer mechanism, students trained without teacher-logit supervision should exhibit substantially weaker activation under a fixed suffix. However, Table 15 reports a triggered anchor probability of 0.98 at λ = 0, compared with 0.87 at λ = 0.5 and 0.85 at λ = 1. I therefore trace the suffix optimization and evaluation calls in the authors’ code. The supplied evaluation script optimizes a separate suffix for each test prompt using the evaluated student’s gradients, then tests it on the same student and prompt. If this procedure was also used for Table 15, the strong activation at λ = 0 could reflect direct white-box suffix optimization rather than inherited teacher preferences. Even if this procedure were applied consistently to all compared methods, tailoring suffixes to the evaluation model and individual test prompts could inflate ASRs relative to the claimed surrogate-to-victim transfer setting. Equal treatment across methods would not establish transfer without victim-model access or generalization to unseen prompts. The authors should clarify which models and prompts were used for suffix optimization, whether suffixes were re-optimized for each λ setting, and whether optimization involved access to the evaluated students.

Questions for authors

See weaknesses, i will raise my score if the authors could provide more evidences that the performance is mainly from the transfer from the teacher.

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

This paper investigates backdoor transfer through clean-data distillation and proposes DefAttack, which decouples preference transfer from suffix activation. The method uses a margin-based loss and length-adaptive weighting to amplify the teacher's preference for target-domain anchor tokens, allowing students to inherit this preference through clean-data distillation. Activation suffixes are then optimized on independently distilled surrogate students and applied to victim students. Cross-architecture experiments and clean/defect cross-transfer experiments examine transfer effectiveness and the interaction between inherited defects and activation suffixes.

Critical strengths
  1. Decoupling backdoor transfer from activation is a valuable and interesting design. DefAttack amplifies the teacher's target-domain preference, transfers it through clean-data distillation, and then optimizes activation suffixes on surrogate students, without requiring a predefined trigger during teacher modification.

  2. The evaluation is broad and includes controlled experiments on the mechanism. It covers four model families, three tasks, and three types of distillation, alongside robustness analyses, ablations, and defense evaluations.

Critical weaknesses
  1. The distinction between amplifying existing defects and injecting preferences is insufficiently supported. Although Section 2.3 defines model defects, the paper does not specify which systematic biases or shortcut tendencies are exploited, or how they are identified before modifying the teacher. The algorithm directly selects anchor tokens and raises their logits. The experiments therefore more directly support the transfer of artificially specified preferences, without sufficiently establishing the mechanism of existing-defect amplification or its added contribution over prior work on bias transfer.

  2. Key explanations are scattered and incomplete, making the method harder to understand and the experiments harder to assess. The main text describes anchor tokens abstractly, while their selection criteria and task-specific choices appear in Appendix B.1, requiring readers to consult the appendix to understand their connection to the target behavior. Although Section 4.1 and Appendix C list the task datasets, they do not clearly specify the data composition and splits for teacher training, surrogate and victim distillation, suffix optimization, and testing. This makes it difficult to verify the unknown-victim-data setting and the reported suffix generalization.

Questions for authors
  1. Which datasets, backdoor targets, teacher/student checkpoints, and poisoning configurations were used in Figure 1, and why are the Gemma-3 teacher's ASRs relatively low?

  2. For each task in the main experiments, how are the datasets for teacher training, surrogate and victim distillation, suffix optimization, and testing constructed and split, and is there any data overlap across these stages?

  3. Why is there no empirical comparison with BIASED-R?

Reviewer_Ta1aOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DefAttack, a backdoor attack that transfers from a poisoned teacher to student LLMs even when knowledge distillation uses only clean data. Instead of relying on a predefined trigger-response association, DefAttack amplifies latent preferences for target-domain anchor tokens in the teacher’s output logits, causing students to inherit the defect through distillation. After distillation, the attacker optimizes a transferable suffix on surrogate students to activate the inherited behavior. Experiments across four model families and three tasks report high attack success rates with low false-positive rates under multiple distillation settings.

Critical strengths
  1. DefAttack increases the harmfulness of backdoor propagation, even when the downstream process relies solely on clean data for distillation.

  2. The paper decouples the propagation'' and activation'' of backdoors, achieving propagation by amplifying the preference for anchor tokens in the teacher’s output distribution, and then utilising post-distillation to optimise the suffix and activate latent defects in the student.

  3. The authors conducted evaluations across four model families, three task categories and various distillation settings, and supplemented these with cross-architecture transfer, ablation analysis, and experiments on concealment and defence.

Critical weaknesses
  1. The research background presented in the abstract is inappropriate. There is already existing foundational work on backdoor transferability in the context of knowledge distillation (KD). It is clear that this paper should primarily focus on addressing the limitations of such prior work; therefore, a more fitting approach would be to identify specific issues within that existing research and propose targeted solutions.

  2. Analysis indicates that this backdoor transmission stems from inherent model flaws—a concept that bears a striking resemblance to the findings in "Language models transmit behavioural traits through hidden signals in data." Since the clean data contains no triggers, what is being transmitted are, in fact, certain intrinsic characteristics of the model itself. This raises a question: does the standard definition of a "backdoor" still hold in this context? Is a specific trigger—an additional input—actually required to activate the backdoor?

  3. Does the architecture of the agent model need to be consistent with that of the student model? This appears to be a black-box adversarial optimisation problem. Furthermore, do the architectures of the teacher model and the student model also need to be consistent? How are differences in vocabulary handled?

  4. The definition of "anchor tokens" requires further clarification.

Questions for authors
  1. Does this work not require poisoned data? Is the teacher model merely amplifying bias? If so, I would argue that this is a form of adversarial attack rather than a backdoor.

  2. If an attack ultimately still requires an optimized suffix to be activated, what new security capability does the "decoupling of propagation and activation" actually provide? Furthermore, how does this suffix differ—in terms of its mechanism and the threat model it operates under—from existing universal adversarial suffixes or jailbreak prompts?

  3. How do attackers get an activation suffix into the input of a victim student model?

  4. Does ASR truly measure the semantic behavior specified by an attacker, rather than merely the presence of anchor tokens or domain-specific target vocabulary? For tasks involving jailbreaks, refusals, and sentiment, are the full behaviors verified through human evaluation or a robust "judge" model—instead of relying on automated metrics that are easily manipulated by keywords?

  5. To what extent does the attack depend on the distillation data aligning with the task domain assumed by the attacker? If the victim uses distillation with data from a different domain, a different language, or a large-scale general-purpose corpus, will the defect preference still be inherited?

Comments on the paper

1 comment

Thank you for the detailed rebuttal and additional experiments. However, two concerns remain:

  1. W1: Amplifying existing defects vs. injecting preferences. The neutral/sensitive verb experiment demonstrates pre-existing lexical sensitivities but does not establish that DefAttack exploits these sensitivities rather than injecting attacker-specified preferences. The cross-transfer experiments confirm the contribution of teacher modification but do not distinguish these mechanisms. Moreover, the new evidence focuses on refusal without establishing analogous pre-existing defects for jailbreak and sentiment. Thus, the central claim of pre-existing defect amplification remains insufficiently supported.

  2. Q1: Gemma-3's lower teacher ASRs. Since the baseline attacks achieve relatively low teacher ASRs on Gemma-3, these results provide limited evidence for assessing backdoor transferability when teacher-side implantation is successful.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.