acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Anti-distillation of Large Language Models: An Unlearnable Examples Perspective

Abstract

Large language models (LLMs) expose valuable capabilities through their generated responses, but making them vulnerable to unauthorized distillation. Existing anti-distillation methods largely assume that knowledge transfer relies on fitting to teacher-generated responses and therefore seek to make those responses difficult to learn. We revisit this premise and ask: *Can a student easily fit protected responses yet fail to acquire the generalizable capability behind them?* Inspired by unlearnable examples (UEs), we decouple data fitting from knowledge transfer by injecting easy-to-fit perturbations into the teacher's reasoning, encouraging the student to learn the surface shortcut while neglecting the underlying capability. Building on this principle, we propose DistilShield, which protects LLMs across the distillation trajectory: it leverages an *Unlearnable Converter* that implants implicit shortcuts into teacher reasoning, steers the learning dynamics away from capability transfer during optimization, and consolidates the induced behavior under the student's own rollout distribution. Extensive experiments demonstrate that DistilShield provides a plug-and-play defense layer that effectively suppresses knowledge distillation while preserving the utility of teacher responses.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Anti-distillation, Large Language Models, Unlearnable Examples, Post-Training

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RVK2OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper examines how to prevent other models from acquiring a language model's capabilities through its generated responses. The authors argue that making responses difficult to learn is not the only way to prevent knowledge transfer. They propose DISTILSHIELD, a framework that modifies the reasoning content of teacher responses to encourage students to learn patterns that do not generalize well. The method uses three main components to introduce these patterns, reduce useful knowledge transfer during training, and maintain their effects when students generate their own responses. The authors also introduce a training algorithm for optimizing the protection method. Experiments across multiple datasets and language models show lower student performance after distillation compared with the tested protection baselines, while maintaining much of the original response utility. More specifically, experiments use two teachers, six students, and six datasets in math, commonsense, and code. With the R1-7B teacher and a Llama 1B student, average student accuracy drops from 38.6 percent with normal distillation to 19.3 percent, below the zero-shot level, while teacher accuracy is unchanged.

Critical strengths

I think that the first strength is the new view of anti-distillation, which separates fitting the data from learning the skill. I think this is an original and useful idea for the area, and the ablations support it well. Moreover, the paper provides extensive experiments involving different models, datasets, and attack settings. The component removal experiments also help explain why the different parts of DISTILSHIELD are needed. Without a doubt, these results strengthen the evidence for the proposed approach.

Critical weaknesses

I noticed that the method requires additional training and response conversion, which may create practical costs for large-scale deployment. Although the paper reports computation times, I think that a clearer study of the cost for a real model service is needed. Moreover, the paper evaluates the protection against a selected set of distillation methods and attempts to overcome the defense. It remains unclear whether more adaptive attackers could identify the modified reasoning patterns, remove them, or recover useful information through other forms of training. This matters because the defense's long-term value depends on its reliability against evolving attack strategies.

Questions for authors

I am curious about the additional computational cost and response delay when using DISTILSHIELD in a real language model service, and how effective the protection would remain if an attacker specifically trained a model to identify and remove the introduced reasoning patterns. Also, please explain how well the defense holds when the attacker uses the preserved teacher answers as RL rewards, or trains only on answers.

2 comments
AuthorsOpenReview ↗

Response to Reviewer RVK2 (Part I)

We sincerely appreciate Reviewer RVK2's recognition of our work and the valuable feedback. We provide a further response to each point below:

Real-time service cost (W1/Q1)

We thank the reviewer for the valuable feedback. Regarding training, the converter is trained only once per dataset, taking about five hours on a single A100 GPU (Table 5). This one-time cost is amortized across all subsequent queries. Regarding serving, our efficiency analysis (Section 5.6 and Table 12) shows that the converter adds only 0.134 s per response, about 20% of the teacher’s own generation time.

To further address the reviewer’s concern, we deploy both the vanilla and the defensive teacher as API services with vLLM and measure additional metrics at different levels of concurrency. To simulate a real-world scenario, we randomly sample 1,000 queries from the six datasets and warm up each service with 50 queries. As shown in the table below, DistilShield increases p50 latency by 3.0-3.8 s (17%-24%) and decreases request throughput by 14%-16%, while token throughput is fully maintained. Importantly, the relative overhead does not increase with concurrency, as the p50 overhead decreases from 24% to 17% as concurrency rises from 1 to 32. Moreover, since the 3B converter incurs a fixed cost independent of the teacher, this overhead is measured against a relatively small 7B teacher and would be marginal for much larger commercial models, which is a desirable property for real-world deployment.

Table: Serving cost under different concurrency levels

ConcurrencyMethodp50 latencyp95 latencyReq/sTokens/sPeak GPU memory
1Vanilla15.7 ± 0.1 s81.6 ± 3.8 s0.044 ± 0.00095 ± 045.8 GB
1DISTILSHIELD19.5 ± 0.2 s84.5 ± 4.3 s0.037 ± 0.00095 ± 071.6 GB
8Vanilla17.2 ± 0.2 s89.9 ± 2.3 s0.321 ± 0.001687 ± 245.8 GB
8DISTILSHIELD20.2 ± 0.1 s93.9 ± 2.2 s0.275 ± 0.001700 ± 472.0 GB
32Vanilla21.8 ± 0.3 s112.1 ± 1.6 s0.961 ± 0.0092066 ± 1145.8 GB
32DISTILSHIELD25.6 ± 0.1 s119.2 ± 1.7 s0.829 ± 0.0042126 ± 1672.0 GB

Protection artifact removal (W2/Q2)

We appreciate the reviewer’s constructive feedback. We further consider two progressively stronger adaptive attacks, which are practical:

  • Sanitization. The attacker attempts to identify and remove protection artifacts (e.g., stereotyped reasoning structures) by rewriting the protected data with another LLM. We implement this by prompting Qwen2.5-7B-Instruct.
  • Privileged sanitization. In an extreme case, the attacker obtains clean versions of some protected responses (e.g., through an insider or a data leak) and uses them to train a dedicated recovery model. We implement this by fine-tuning Qwen2.5-7B-Instruct on the leaked clean-protected pairs.

As shown in the table below, standard sanitization backfires on the attacker, degrading the vanilla-distilled student by 19.4 points and the defensive student even further to 21.5%. Therefore, generic rewriting (e.g., paraphrasing or standard sanitization) cannot distinguish the implanted shortcut from the useful content. Privileged sanitization can partially recover 2.4 points of capacity from the defensive teacher, despite requiring rarely available clean responses. These results indicate that the protection is not tied to easily removable artifacts.

Table: Stronger adaptive attacks on GSM8K

Model (Qwen3-8B → Qwen3-0.6B)NoneParaphraseSanitizationPrivileged Sanitization
Vanilla61.357.441.961.9
Defensive (DistilShield)55.952.121.558.3
AuthorsOpenReview ↗

Response to Reviewer RVK2 (Part II)

Only using answers as training signal (Q3)

An attacker can indeed discard the rewritten reasoning and train only on question-answer pairs. However, this also discards the primary supervision signal in reasoning distillation. The setting then is degraded to answer-only fine-tuning or RL, which falls outside the LLM knowledge distillation setting that our work targets. For the same reason, prior anti-distillation works [1,2] do not consider this setting. For completeness, we report the results below. Answer-only SFT performs poorly (12.7% on average), even below distillation from the protected responses. Answer-only RL performs better but still falls 6.5 (RLVR) and 7.7 (GRPO) below distillation from the vanilla teacher with reasoning on average, yielding little gain on MBPP. This confirms that bypassing the defense by discarding the reasoning causes an inevitable cost to the attacker, whether the answers are used for fine-tuning or as RL rewards. More importantly, the answers serve only as correctness labels, and such labels are often obtainable without a teacher (e.g., from public datasets or human annotations).

Table: Distillation with different data components and methods

SettingGSM8KMathMBPP
Clean (reasoning + answer)47.931.130.9
Protected (reasoning + answer)21.116.520.1
Answer only w/ SKD7.210.120.8
Answer only w/ RLVR42.626.621.2
Answer only w/ GRPO41.024.321.4

We hope our response addresses the reviewer’s concerns, and we are glad to engage in any further discussion!

[1] Savani, Yash, et al. "Antidistillation sampling." Advances in Neural Information Processing Systems 38 (2026): 117800-117827

[2] Li, Pingzhi, et al. "Doge: Defensive output generation for llm protection against knowledge distillation." arXiv preprint arXiv:2505.19504 (2025)

Reviewer_Nz64OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes a new perspective on anti-distillation, which utilizes unlearnable examples to prevent distilled datasets from improving student models' performance. Specifically, the proposed method, DistillShield, uses an unlearnable converter to implant implicit shortcuts into distilled samples. Furthermore, to optimize the unlearnable converter, this work proposes Anti-GDPO, which incorporates four distinct rewards. Experiments on diverse datasets show that DistillShield can preserve the teacher model's performance while significantly degrading the learning performance of student models.

Critical strengths
  1. The topic discussed in this work is critical, since anti-distillation is an important way to protect commercial property.
  2. The presented figures are well-designed. However, the content of these figures is partially unclear.
  3. This work conducts extensive experiments to show the superiority of the proposed method.
Critical weaknesses
  1. Although this work claims to revisit the anti-distillation and to be the first to study this problem from the unlearnable example perspective, the core of the proposed method isn't too different from the former "making those responses difficult for students to fit" mentioned in the paper. Papers such as [1] also make the distilled datasets unlearnable while preserving the teacher models' performance.

  2. In Figure 1A, the definitions and metrics of "student accuracy" and "generalizability" are unclear. For instance, which datasets are used in Figure 1A? Furthermore, why can student accuracy represent generalizability?

  3. Additionally, the motivation behind Anti-GDPO is unclear. The paper does not clarify why Anti-GDPO is necessary, nor why alternative methods fail to optimize the unlearnable converter.

  4. In practice, most unauthorized distillations rely on commercial APIs. However, this work does not evaluate or discuss the performance of the proposed method on commercial APIs.

[1] Antidistillation Sampling. NIPS 2026.

Questions for authors

see weakness

6 comments
AuthorsOpenReview ↗

Response to Reviewer Nz64 (Part I)

We sincerely appreciate the reviewer’s critical feedback. We provide further explanations and responses for each point below:

Distinction with previous work (W1)

Distinction. We thank the reviewer for the critical question. We respectfully clarify that DistilShield optimizes in the opposite direction from ADS [1] and prior defenses. ADS perturbs the teacher’s next-token distribution toward tokens that increase a proxy student’s downstream loss. This is an error-maximizing (min-max) objective (a.k.a. poisoning/adversarial objective), min⁡θ  max⁡y~  L(θ;y~)\min_{\theta} \; \max_{\tilde{y}} \; \mathcal{L}(\theta; \tilde{y}), which makes the data harder to learn from. In contrast, DistilShield builds on an error-minimizing (min-min), min⁡θ  min⁡y~  L(θ;y~)\min_{\theta} \; \min_{\tilde{y}} \; \mathcal{L}(\theta; \tilde{y}). Our converter explicitly minimizes the student’s loss on the protected responses (Eq. 5), intentionally making them easy to fit. Shortcut binding (Eq. 8) then steers the resulting updates away from the direction of clean knowledge transfer. The key idea is therefore not to prevent fitting, but to let the student fit a shortcut that does not transfer the teacher’s underlying capability. Our results support this distinction:

  • Table 6: DistilShield responses have lower perplexity than the original teacher responses (6.1 vs. 9.6), meaning they are more predictable and easier to fit.
  • The case studies (Figure 4 and Appendix J): the converter transforms the reasoning into a stereotyped template that makes it easy for students to learn.

Further evidence. We respectfully invite the reviewer to refer to the unlearnable examples literature [2], which discusses more detailed distinctions. To make it concrete in the discussion, we also plot the training dynamics of students distilled from clean, DistilShield, and ADS responses in Figure 5 of Appendix G.10 of the revised manuscript, with a condensed version below. As shown in the table, the two defenses exhibit opposite training dynamics relative to clean distillation. ADS consistently yields a higher training loss (1.25 vs. 0.87 at the final step) and exhibits less stable optimization, with more pronounced loss spikes, as more clearly shown in Figure 5. In contrast, DistilShield consistently maintains a lower training loss (0.22 vs. 0.87). This confirms that ADS protects by making responses harder to fit, whereas DistilShield exploits easy-to-fit shortcuts.

Table: Training dynamics comparison on CSQA dataset

Metric/Step021426384105126147
Clean: Training Loss2.53071.83261.16220.88430.95840.96221.04770.8675
DistilShield: Training Loss1.25820.61690.45250.68150.35910.29470.32010.2221
ADS: Training Loss3.01712.06531.64051.18711.39771.14821.31071.2513

Utility trade-off. Moreover, ADS does not fully preserve the teacher’s performance: it achieves protection at the cost of teacher accuracy. As shown in Table 1, ADS reduces the teacher’s average accuracy from 80.2% to 59.3%, with drops to 26.0% on HumanEval and 32.9% on MBPP. In contrast, DistilShield preserves teacher accuracy.

Explanation of Figure 1A (W2)

We appreciate the reviewer’s careful observation and clarify Figure 1A below.

Setup. Figure 1A uses GSM8K, with DeepSeek-R1-Distill-Qwen-7B as the teacher and Llama-3.2-1B as the student. A human-designed shortcut pattern is injected into 1,000 randomly sampled teacher reasoning traces. The left axis shows the student’s training loss on this shortcut-injected distillation corpus, and the right axis shows the student’s accuracy on the held-out test set. The dashed line marks the student’s zero-shot accuracy. We have added these details in Appendix F.3 of the revised manuscript.

Explanations for generalization. The distillation queries come from the training split, and evaluation uses the disjoint test split (Appendix F.1). The test questions are clean and contain no injected shortcut, and the student generates answers autoregressively on its own. Accuracy on these unseen test problems therefore measures the transferable problem-solving capability the student acquired. This follows the standard practice for measuring generalization in both anti-distillation [1] and unlearnable examples [2].

AuthorsOpenReview ↗

Response to Reviewer Nz64 (Part II)

Explanation of Anti-GDPO (W3)

We thank the reviewer for the thoughtful question. We would like to kindly note that the paper discusses this motivation in Section 4.4 and Appendix C (lines 1029-1042), and we summarize it below.

Why RL. The overall objective (Eq. 12) involves discrete sampling of rewrites and nested optimization of the proxy student (lines 264-265), so gradients cannot be backpropagated directly through the converter. RL tackles this by treating each objective as a reward on a detached simulated trajectory.

Why GDPO. Our four reward channels have heterogeneous scales and variances: a negative log-likelihood, a normalized inner product, a KL divergence, and a bounded similarity score. Standard GRPO sums them before normalization, which allows the largest or noisiest channel to dominate the advantage (lines 270–272). Following GDPO, Anti-GDPO standardizes each channel independently before aggregation to mitigate this.

Empirical evidence. The ablation in Table 3 isolates this design choice by replacing Anti-GDPO with GRPO using the same rollouts and rewards. This weakens protection by 1.8 points on average across the six datasets (Table 9), confirming the benefit of Anti-GDPO.

Commercial models (W4)

We thank the reviewer for the valuable feedback. We would like to respectfully point out that in the anti-distillation setting, the defender is the model provider and thus has full access to its own serving pipeline, including the original teacher outputs. To faithfully reflect this defender assumption, we therefore conduct experiments on open-weight models rather than commercial models, same as prior work [1,2]. To improve comprehensiveness, we further evaluate DistilShield on three state-of-the-art commercial teachers. As shown in the table below, DistilShield consistently reduces the accuracy of the distilled student across all teachers and both tasks, by 7.2 points on average and by up to 9.3 points (Gemini 3.7 Flash on HumanEval). This demonstrates effective protection for commercial systems. Moreover, in real-world deployment, commercial model developers have full access to their own models and pipelines, enabling even more effective protections.

Table: Commercial teacher models

Teacher (Student: Llama-3.2-1B)GSM8K: GPT-6.1 SolGSM8K: Claude Sonnet 5.5GSM8K: Gemini 3.7 FlashHEval: GPT-6.1 SolHEval: Claude Sonnet 5.5HEval: Gemini 3.7 Flash
Vanilla41.348.851.230.026.034.0
Defensive (DistilShield)33.640.045.824.719.324.7

We hope our response addresses the reviewer’s concerns, and we are glad to engage in any further discussion!

[1] Savani, Yash, et al. "Antidistillation sampling." Advances in Neural Information Processing Systems 38 (2026): 117800-117827

[2] Huang, Hanxun, et al. "Unlearnable examples: Making personal data unexploitable." ICLR (2021).

[3] Li, Pingzhi, et al. "Doge: Defensive output generation for llm protection against knowledge distillation." arXiv preprint arXiv:2505.19504 (2025)

Reviewer_Nz64OpenReview ↗

Replies for W4

Thanks for your detailed and clear replies.

I have a confusion about your experiment pipeline on commercial models.

Since your method needs the reasoning traces of teacher model, however, the explicit reasoning traces of Claude Sonnet 5.5 and GPT-6.1 Sol are unaccessible. Could your further clarify that?

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

In this paper, it revisits anti-distillation for large language models from the perspective of unlearnable examples. Instead of making teacher-generated responses difficult to fit, the authors investigate whether a student can fit protected responses while failing to acquire the teacher's generalizable capabilities. Based on this idea, they propose DISTILSHIELD, which uses an Unlearnable Converter to rewrite reasoning traces while preserving the final answers. The framework combines three components: shortcut induction, gradient-based shortcut binding, and on-policy shortcut consolidation, optimized through the proposed ANTI-GDPO algorithm. Experiments on six datasets covering mathematical reasoning, commonsense reasoning, and code generation show that DISTILSHIELD generally provides stronger anti-distillation protection than existing baselines while maintaining teacher answer accuracy. The paper also includes ablation studies, cross-model evaluations, robustness tests, and response-quality assessments to support its findings.

Critical strengths
  1. Most existing approaches focus on making teacher responses harder to imitate, whereas this paper asks whether the student can learn the surface patterns without acquiring the underlying reasoning capability. This is an interesting way to rethink the problem, and I believe it provides a meaningful conceptual contribution beyond simply modifying existing distillation objectives.

  2. Another strength is that DISTILSHIELD addresses several challenges that arise when applying unlearnable examples to pretrained LLMs. The combination of shortcut induction, gradient anti-alignment, and on-policy consolidation is reasonably motivated, especially given that pretrained students may already possess relevant knowledge. I also appreciate that the authors connect the gradient alignment objective to the first-order change in clean distillation loss, which helps explain the design.

  3. What I find useful here is that the defense only modifies the reasoning trace rather than the final answer. This creates a clear separation between protecting the teacher's knowledge and maintaining its answer-level utility. Table 1 shows that DISTILSHIELD preserves the teacher's measured accuracy, whereas some competing methods substantially reduce it. This makes the proposed approach more appealing for practical deployment.

Critical weaknesses
  1. While the overall results are encouraging, I noticed that the improvements vary considerably across model architectures. For example, Table 2 shows relatively modest reductions for some Qwen and Gemma students compared with the much larger improvement for Llama-3.2-1B on GSM8K. A few code-generation results also show only small differences from competing methods.

  2. One concern I have is that the robustness experiments consider a fairly limited set of countermeasures. Section 5.5 examines adversarial training, repetitive querying, and paraphrasing, which are reasonable starting points. However, the repetitive-querying experiment uses only two queries per training question, and the paraphrasing strategy relies on a single rewriting procedure. An attacker might instead remove reasoning traces, use more aggressive filtering, or combine protected responses with additional clean training data.

  3. Another limitation is that DISTILSHIELD assumes the teacher response can be separated into a reasoning trace and a final answer. This works naturally for the reasoning-oriented tasks considered in the paper, but the applicability to models exposing only final answers, structured outputs, or very short explanations is less clear. In addition, the experiments involve two relatively small teacher models, so the extension to much larger commercial systems remains uncertain.

  4. I found some inconsistencies that should be straightforward to address. For instance, Table 1 reports 21.1% student accuracy on GSM8K for DISTILSHIELD, whereas Appendix Table 10 reports 33.0% for batch size 1, which is also the default configuration in Appendix F.3. Appendix Table 11 further reports 44.1% for the defensive teacher under the "None" condition for the same teacher-student pairing. It is unclear whether these differences arise from distinct experimental protocols or reporting errors. There also appears to be a notation inconsistency between Equations 4 and 12 regarding which part of the response is rewritten.

Questions for authors
  1. Could the authors explain why DISTILSHIELD provides substantially stronger protection for some student architectures than others? In particular, is the variation mainly related to model capacity, pretrained knowledge, tokenizer differences, or the compatibility between the proxy student and the actual target student?

  2. Have the authors evaluated an attacker that simply removes the rewritten reasoning traces and distills from question-answer pairs, or combines protected responses with additional clean data? It would be helpful to understand whether the protection depends on the attacker retaining the modified reasoning during training.

  3. Could the authors provide additional evidence that the proposed objectives specifically suppress knowledge transfer rather than merely making fine-tuning less effective? For example, comparing the protected-response training loss, clean-response training loss, and held-out accuracy throughout the distillation process would help clarify the mechanism.

2 comments
AuthorsOpenReview ↗

Response to Reviewer WMDv (Part I)

We sincerely appreciate Reviewer WMDv's recognition of our work and the constructive feedback. We provide a further response to each point below:

Protection variation across student architectures (W1/Q1)

We appreciate the reviewer’s insightful observation. We believe the variation mainly reflects the discrepancy in learning dynamics between the proxy and the target student. Differences in architecture, capacity, and pretrained knowledge affect how reliably a shortcut optimized on the proxy induces similar learning behavior in the target. Consistent with this, Table 2 demonstrates that larger students with stronger prior knowledge (e.g., Qwen3-1.7B) retain higher accuracy. Tokenizer differences, by contrast, do not appear to be the dominant factor. Llama-3.2-1B uses a different tokenizer from our Qwen2.5-3B proxy yet receives the strongest protection, whereas Qwen2.5-0.5B shares the proxy’s tokenizer family but receives more modest protection. The parameter sensitivity experiment shown in Figure 3 supports this interpretation. By varying the proxy while holding the target student model fixed, the protection also changes.

Additionally, we invite the reviewer to note that the relatively modest improvement is mainly observed in HEval. This aligns with intuition, since HumanEval provides only 114 distillation queries, compared with about 7.5K for GSM8K. This much weaker supervision signal limits the margin between all methods.

Data mixing, reasoning removal and more adaptive attacks (W2/W3/Q2)

We thank the reviewer for the constructive comments. We discuss the suggested settings one by one below.

Data mixing. We create a dataset comprising a mixture of clean and protected data and then vary the mix ratio. As shown in the table below, student accuracy decreases steadily as the proportion of protected data grows, from 47.9% to 21.1% on GSM8K and from 31.1% to 16.5% on MATH. With only 40% of the protected data, accuracy drops by 10.0 and 8.0 points, respectively. This indicates that the shortcut is not easily diluted by clean data, and even a portion of protected responses substantially suppresses knowledge transfer.

Table: Data mixing

Dataset (R1-7B → Llama-3.2-1B)020406080100
GSM8K47.941.537.934.230.721.1
Math31.127.423.121.019.916.5

Reasoning removal. An attacker can indeed discard the rewritten reasoning and train only on question-answer pairs. However, this also discards the primary supervision signal in reasoning distillation. The setting is then degraded to answer-only fine-tuning, which falls outside the LLM knowledge distillation setting that our work targets. For the same reason, prior anti-distillation works [1,2] do not consider this setting. For completeness, we report the results below. Answer-only training performs poorly, falling 23.9 points below distillation from the vanilla teacher on average. It is even worse than distilling from the protected responses on GSM8K and MATH (7.2% vs. 21.1% and 10.1% vs. 16.5%), and comparable on MBPP (20.8% vs. 20.1%). This confirms that discarding the reasoning costs the attacker much more transferable capability than it recovers.

Table: Distillation with different components

Teacher (R1-7B → Llama-3.2-1B)GSM8KMathMBPP
Vanilla (reasoning + answer)47.931.130.9
Defensive (reasoning + answer)21.116.520.1
Answer only7.210.120.8

Stronger adaptive attacks. Inspired by the reviewers’ suggestion, we further evaluate two progressively stronger adaptive attacks, which we also believe are practical:

  • Sanitization. The attacker attempts to identify and remove protection artifacts (e.g., stereotyped reasoning structures) by rewriting the protected data with another LLM. We implement this by prompting Qwen2.5-7B-Instruct.
  • Privileged sanitization. In an extreme case, the attacker obtains clean versions of some protected responses (e.g., through an insider or a data leak) and uses them to train a dedicated recovery model. We implement this by fine-tuning Qwen2.5-7B-Instruct on the leaked clean-protected pairs.

As shown in the table below, standard sanitization backfires on the attacker, degrading the vanilla-distilled student by 19.4 points and the defensive student even further to 21.5%. Therefore, generic rewriting cannot distinguish the implanted shortcut from the useful content. Privileged sanitization can partially recover 2.4 points of capacity from the defensive teacher, despite requiring rarely available clean responses. These results indicate that the protection is not tied to easily removable artifacts.

Table: Stronger adaptive attacks on GSM8K

Model (Qwen3-8B → Qwen3-0.6B)NoneParaphraseSanitizationPrivileged Sanitization
Vanilla61.357.441.961.9
Defensive (DistilShield)55.952.121.558.3
AuthorsOpenReview ↗

Response to Reviewer WMDv (Part II)

Commercial models (W3)

We thank the reviewer for the valuable feedback. We would like to respectfully point out that in the anti-distillation setting, the defender is the model provider and thus has full access to its own serving pipeline, including the original teacher outputs. To faithfully reflect this defender assumption, we therefore conduct experiments on open-weight models rather than commercial models, same as prior work [1,2]. To improve comprehensiveness, we further evaluate DistilShield on three state-of-the-art commercial teachers. As shown in the table below, DistilShield consistently reduces the accuracy of the distilled student across all teachers and both tasks, by 7.2 points on average and by up to 9.3 points (Gemini 3.7 Flash on HumanEval). This demonstrates effective protection for commercial systems. Moreover, in real-world deployment, commercial model developers have full access to their own models and pipelines, enabling even more effective protections.

Table: Commercial teacher models

Teacher (Student: Llama-3.2-1B)GSM8K: GPT-6.1 SolGSM8K: Claude Sonnet 5.5GSM8K: Gemini 3.7 FlashHEval: GPT-6.1 SolHEval: Claude Sonnet 5.5HEval: Gemini 3.7 Flash
Vanilla41.348.851.230.026.034.0
Defensive (DistilShield)33.640.045.824.719.324.7

Knowledge transfer vs ineffective training (Q3)

We appreciate the reviewer’s important suggestion. We track the training loss and held-out accuracy throughout distillation for students trained on both clean and DistilShield-protected data. The full curves are shown in Figure 5 of Appendix G.10 of the revised manuscript, and a condensed version is presented in the table below. The student fits DistilShield responses even more easily than clean ones, with training loss consistently lower, reaching 0.22 vs. 0.87. However, its held-out accuracy drops from 54.0% to 18.8%, whereas clean distillation improves accuracy to 57.4%. Notably, accuracy declines sharply as training loss decreases, indicating that the student learns the implanted shortcut rather than the transferable capability. This confirms that DistilShield decouples data fitting from knowledge transfer.

Table: Training dynamics on CSQA dataset

Metric/Step021426384105126147
Clean: Training Loss2.53071.83261.16220.88430.95840.96221.04770.8675
Clean: Eval Acc54.054.153.754.556.357.057.357.4
DistilShield: Training Loss1.25820.61690.45250.68150.35910.29470.32010.2221
DistilShield: Eval Acc54.043.229.530.128.819.320.018.8

Different results and typo (W4)

We thank the reviewer for the careful observation. The parameter sensitivity experiments in Table 10 use 1,000 randomly sampled training questions as distillation queries for efficient evaluation across various configurations, which is different from the whole training set used in the main experiments. We have also removed the results using legacy settings in Table 11 to avoid confusion and corrected symbol typos in the revised manuscript.

We hope our response addresses the reviewer’s concerns, and we are glad to engage in any further discussion!

[1] Savani, Yash, et al. "Antidistillation sampling." Advances in Neural Information Processing Systems 38 (2026): 117800-117827

[2] Li, Pingzhi, et al. "Doge: Defensive output generation for llm protection against knowledge distillation." arXiv preprint arXiv:2505.19504 (2025)

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We thank all reviewers for their insightful and constructive reviews of our manuscript. We are encouraged to hear that the reviewers found our perspective novel and conceptually meaningful (23\frac{2}{3} Reviewers WMDv, RVK2), our method well-motivated and well supported by ablations (23\frac{2}{3} Reviewers WMDv, RVK2), and our defense practical since it preserves teacher utility (23\frac{2}{3} Reviewers WMDv, RVK2). They also found our experiments extensive (33\frac{3}{3} Reviewers WMDv, Nz64, RVK2) and our presentation clear with well-designed figures (23\frac{2}{3} Reviewers WMDv, Nz64). Below, we respond to each reviewer's concerns point by point, and we have uploaded a revised manuscript. The main changes are highlighted in blue and appended to the end of the corresponding sections, so the original section number is largely preserved for convenience of discussion:

I. Stronger adaptive attacks: As suggested by Reviewers WMDv and RVK2, we evaluate three additional families of countermeasures:

  • Data mixing: protected responses are mixed with clean data at ratios from 0% to 100% (Appendix G.7).
  • Answer-only distillation: the attacker discards the rewritten reasoning and trains on question-answer pairs via SKD, RLVR, and GRPO (Appendix G.8).
  • Artifact removal: the attacker sanitizes the protected responses with an LLM (Qwen2.5-7B-Instruct). In an extreme privileged setting, the attacker instead trains a dedicated recovery model on leaked clean-protected pairs (Appendix G.4).

II. Commercial teacher models: As suggested by Reviewers WMDv and Nz64, we evaluate DistilShield on three state-of-the-art commercial models (GPT-6.1 Sol, Claude Sonnet 5.5, and Gemini 3.7 Flash) on GSM8K and HumanEval (Appendix G.9).

III. Mechanism analysis via training dynamics: As suggested by Reviewers WMDv and Nz64, we track three quantities throughout distillation: the training loss on protected responses, the training loss on clean responses, and held-out accuracy. We further compare these dynamics with ADS [1] (Appendix G.10).

IV. Real-world serving cost: As suggested by Reviewer RVK2, we deploy the vanilla and defensive teachers as API services with vLLM. On 1,000 queries sampled from all six datasets, we measure p50/p95 latency, throughput (requests/s and tokens/s), and peak GPU memory at concurrency levels of 1, 8, and 32 (Appendix G.11).

V. Detailed clarifications & settings: We clarify the experimental settings that may have caused confusion, to improve clarity and reproducibility (Appendix F.3):

  • Clarification of Figure 1A: As suggested by Reviewer Nz64, we detail the setup of the warm-up study.
  • Parameter sensitivity: As suggested by Reviewer WMDv, we clarify that the parameter sensitivity experiments use 1,000 randomly sampled data points.

VI. Minor issues: As suggested by Reviewer WMDv, we correct the typos and other minor flaws.

At the same time, we would like to respectfully note that prior anti-distillation works [1,2,3] did not evaluate stronger adaptive attacks (I), commercial teacher models (II), or real-world serving costs (IV), as some of those go beyond established settings (e.g., out of the distillation scope or compromise the defender's assumptions). We nevertheless include them in the rebuttal to more rigorously assess the robustness, generality, and practicality of DistilShield.

We thank all reviewers again for the time and valuable efforts in helping us improve the manuscript. We hope our responses have addressed all concerns, and we would sincerely appreciate it if you could take them into account in the final assessment. We are happy to engage in any further discussion!

[1] Savani, Yash, et al. "Antidistillation sampling." Advances in Neural Information Processing Systems 38 (2026): 117800-117827

[2] Li, Pingzhi, et al. "Doge: Defensive output generation for llm protection against knowledge distillation." arXiv preprint arXiv:2505.19504 (2025)

[3] Ma, Xinhang, et al. "Protecting language models against unauthorized distillation through trace rewriting." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.