acceptodds
Under review as a conference paper at ICLR 2027

Lossless Anti-Distillation Sampling

11 views

Abstract

Frontier commercial generative models face a growing threat from distillation, whereby a distiller harvests generated responses and trains a competing model at drastically lower cost. Existing defenses either modify the generation to degrade distillation performance, sacrificing response quality, or rely on behavioral detection mechanisms that can be readily bypassed through multi-account querying. In this work, we propose Lossless Anti-Distillation Sampling (LADS), which leaves the generation itself unchanged while substantially reducing the effectiveness of distillation. Concretely, LADS uses a coupling mechanism to control the latent randomness, preserving independence within each account while introducing dependence across accounts. By construction, each benign user, who typically holds only a single account, receives independent samples from the original model across visits, thereby enjoying a lossless experience. However, for a multi-account task-specific distiller, semantically similar queries submitted across different accounts are assigned coupled randomness, inducing dependence in the harvested data and thereby degrading the generalization performance of the distilled model. Using uniform convergence theory, we show that LADS provably degrades the distiller’s generalization gap relative to standard i.i.d. sampling. Experiments on image generation, mathematical reasoning, and code generation confirm that LADS substantially degrades the performance of distilled students while preserving exact statistical fidelity for individual users.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
anti-distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_uD7TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LADS, a defense against multi-account model distillation. The main idea is to couple sampling randomness across accounts while preserving the original output distribution for each individual account. The authors provide a theoretical analysis and evaluate the method on image generation, mathematical reasoning, and code generation.

Critical strengths
  1. I liked the main idea of controlling sampling randomness rather than modifying model outputs. Preserving the single-account user experience while reducing the effectiveness of multi-account distillation is an interesting direction.

  2. The experiments show promising results across both diffusion and language models. The theoretical analysis also provides a useful perspective on how correlated training samples affect effective sample size.

Critical weaknesses
  1. My main concern is the theoretical guarantee. Theorem 3.2 appears to rely on balanced center-depth weights, as clarified in Appendix C.8, but this condition is not stated explicitly in the main theorem. Also, a weaker generalization upper bound does not necessarily establish that the student's actual generalization error must increase.

  2. The effectiveness against adaptive distillers remains unclear. As the authors acknowledge in Section 3.3, an attacker could modify queries to reach different semantic buckets or change query allocation across accounts. The experiments mainly consider noisy bucket assignments rather than these adaptive strategies. This limits the practical security conclusions.

Questions for authors
  1. Could the authors clarify the balanced-weight assumption in Theorem 3.2? Does the claimed effective sample size reduction still hold when query frequencies across buckets are highly uneven?

  2. How robust is LADS when the distiller deliberately distributes queries across different semantic buckets or changes the query order across accounts?

  3. In Table 2, the reported improvements differ from the numbers discussed in Section 4.2. Could the authors correct these values and clarify the actual performance reductions?

2 comments

We thank the reviewer for the careful reading. Distillation is increasingly recognized as a major threat to commercial LLM providers. Anthropic was even speculated to have downgraded Claude’s response quality to mitigate distillation attacks, drawing substantial public criticism. As the first lossless anti-distillation method, we believe our work opens a timely research direction, and we hope the reviewers will take its exploratory nature and potential impact into consideration.

Weakness 1: Interpretation of theory

Thanks for raising the question. A central contribution of generalization theory is to provide “high-probability guarantees” for learning tasks. In our case, standard i.i.d. sampling provides a stronger high-probability guarantee on test loss than LADS, and the bound is tight up to constant factors. See lines 229–231 of the paper. Such comparisons of statistical guarantees are standard and meaningful in learning theory, although they do not imply that “the student’s actual generalization error must increase” for every distribution. This is why experiments are still necessary. Our results show that LADS consistently reduces distillation effectiveness across both image and language generation tasks, providing direct empirical evidence of its practical impact.

Question 1: Technical issues

Thanks for raising this question. The balanced case in Theorem 3.2 gives a concise form of the general weighted analysis in Appendix C.7–C.8. Under clean assignments, uniform weights maximize neffn_{\mathrm{eff}} for a fixed set of cells; the general bound tracks the actual weights when query frequencies are uneven.

The source of sample compression is the reuse of randomness across accounts, and the weighted analysis measures how often each draw is reused. We will state the balancing condition explicitly and clarify this connection in the paper.

Question 2: Robustness

We did consider how to evaluate the robustness of our algorithm and explicitly discuss the challenges of experimental design in the paper (lines 349–351). A key practical consideration is that distillers typically lack full knowledge of the provider’s sampling strategy, including the exact bucket assignments and implementation details. Recovering this information through reverse engineering can also be challenging. We therefore incorporate noisy semantic mappings into our experiments as one way to model this uncertainty. This experimental design is also intended to address the concern raised in Weakness 2.

The adversarial strategy described in Question 2 does not eliminate the cross-account coupling in LADS. Under Algorithm 2, queries from different accounts share randomness whenever their bucket assignments and within-bucket visit counts match. Reordering queries changes which queries are coupled, but preserves this mechanism.

For example, consider two accounts: account 1 submits queries q11q_{11} and q12q_{12}, while account 2 submits q21q_{21} and q22q_{22}. Suppose all four queries fall into the same bucket. If both accounts submit their queries in the original order, q11q_{11} shares randomness with q21q_{21}, and q12q_{12} shares randomness with q22q_{22}. If account 1 reverses its query order, q12q_{12} instead shares randomness with q21q_{21}, and q11q_{11} shares randomness with q22q_{22}.

In both cases, the four responses are generated using only two independent randomness realizations. Thus, query reordering preserves the reduction in independent randomness underlying the collected responses.

To be honest, we have put considerable effort into designing experiments that capture realistic attack–defense interactions in model distillation. We would welcome further suggestions on the experimental design and would be happy to incorporate them into our evaluation.

Question 3: Numbers mismatch

Thank you for pointing out this inconsistency. The numbers in Table 2 are correct, but we forgot to update the corresponding values in Section 4.2. We will revise the text to match the table.

Dear reviewer,

Thank you for your constructive review and comments! We have provided point-to-point responses to your review. Considering the author-reviewer discussion is ending very soon, could you kindly take a look at our responses and let us know your feedback? We are happy to discuss more if you still have any concerns. Thank you again and look forward to your feedback.

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

This paper proposes Lossless Anti-Distillation Sampling (LADS), a method to protect generative models against multi-account model distillation without modifying the output distribution observed by individual users. The key idea is to couple the sampling randomness across different accounts based on semantic query buckets and per-account access counts, introducing statistical dependence into the collected training data while preserving within-account sampling behavior.

The authors provide a theoretical analysis using uniform convergence bounds to characterize the reduction in effective sample size. Experiments on diffusion-based image generation and language model distillation for mathematical reasoning and code generation demonstrate degraded student performance under the proposed sampling strategy, while maintaining the original marginal generation distribution for individual users.

Critical strengths

The paper introduces a method anti-distillation by manipulating cross-account sampling dependence rather than directly perturbing model outputs. Preserving the original generation distribution while reducing the statistical efficiencently. The paper also evaluate the method on both diffusion-based image generation and autoregressive language models, including mathematical reasoning and code generation tasks. The consistent degradation in student performance across these settings provides useful empirical evidence for the proposed mechanism.

Critical weaknesses

A minor concern that Theorem 3.2 relies on restrictive assumptions. These conditions are not guaranteed for a practical or adaptive distillation adversary. May substantially limits the generality of the theoretical claims.

Questions for authors

See weakness

2 comments

We thank the reviewer for their support of this work. Distillation poses a growing threat to commercial LLM providers, and our goal is to develop a practical defense. Our theoretical analysis aims to provide conditions under which LADS reduces the effective number of independent samples and characterizes the resulting generalization guarantees.

Like most theoretical analyses, ours relies on simplifying assumptions that may not fully capture real-world conditions. To evaluate its practical effectiveness, we carefully considered how our experiments could better reflect real-world usage, where users typically lack full knowledge of the provider’s sampling strategy, including the exact bucket assignments and implementation details. We therefore incorporate noisy semantic mappings into our experiments as one way to capture this uncertainty. We would also welcome the opportunity to explore additional realistic settings in future work.

Dear reviewer,

Thank you for your constructive review and comments! We have provided point-to-point responses to your review. Considering the author-reviewer discussion is ending very soon, could you kindly take a look at our responses and let us know your feedback? We are happy to discuss more if you still have any concerns. Thank you again and look forward to your feedback.

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

Summary: The paper develops a sampling method to reduce efficacy in learning from already trained models. The premise is that for a model to be trained adequately the student has to send queries from multiple accounts otherwise they would get flagged before that. Given this premise they reduce exploration of the randomness by fixing the random seed of the teacher, when generating answers across similar prompts in the query space. The reported upper without LADS (their sampling technique) degrades from (1/\sqrt{KT}) to (1/\sqrt{NT}), where N is the number of dissimilar query regions.

I lean towards rejection at this point.

Critical strengths

Strengths: This is an interesting and timely problem. Some current attempts are tailored for single account attacks, and their results touch on the realm where the attackers are forced to change accounts when they are unable to collect enough teacher responses with one account.

Critical weaknesses

Weakness: The authors establish only a worse upper for their LADS model.

Weakness: At lines 1040 - 60 After conditioning on the shared random hash h, Hoeffding concentrates the indicators around rho(h), in each slice, however we do not have independence for the indicators unconditional to h. The proof appears to substitute the bound on Eh[ρ(h)]\mathbb E_h[\rho(h)] for a bound on rho(h) itself.

Questions for authors

Question: Can this bound be shown to be tight?

Question: Can the authors justify that substitution, or give a revised high-probability bound over the hash that still yields the claimed result?

4 comments

We appreciate the reviewer’s concerns about the theoretical assumptions. We would also like to note that the review focuses on these assumptions without discussing the practical motivation or experimental results of the paper. Our work addresses the practical problem of multi-account distillation through a combination of problem formulation, algorithm design, theoretical analysis, and empirical evaluation. The theoretical and empirical results provide complementary evidence for the proposed approach. We would therefore appreciate consideration of the algorithmic contributions and experimental findings alongside the concerns about the theoretical assumptions in the overall assessment of the paper.

Weakness 1

We do consider the tightness of the bounds and use three references to explicitly discuss it in lines 229–231 of the paper. Furthermore, we would like to emphasize that the key point of LADS is the exact reduction in the number of independent samples. This loss of independent information provides a statistical basis for expecting poorer generalization. Crucially, the reduction itself is a structural property of the sampling scheme and holds independently of the particular concentration inequality employed.

Weakness 2

Thanks for pointing this out. We will revise the concentration step by separating hash selection from query sampling and making the resulting probability guarantee explicit. This part of the analysis quantifies how reliably queries are assigned to their intended semantic buckets. The general bound in Appendix C.7 already accounts for imperfect assignments. This is also why we evaluate LADS with noisy semantic mappings: the experiments show that cross-account coupling can hinder distillation across both image and language generation even when the grouping is imperfect.

Dear reviewer,

Thank you for your constructive review and comments! We have provided point-to-point responses to your review. Considering the author-reviewer discussion is ending very soon, could you kindly take a look at our responses and let us know your feedback? We are happy to discuss more if you still have any concerns. Thank you again and look forward to your feedback.

Thank you for addressing my concerns. I will update my score to 3.

We’re glad to hear that our response addressed your concerns. Thank you for your support!

Comments on the paper

2 comments

Dear Reviewers,

Authors’ kindly responded to address your concerns. Please check rebuttals and share your opinions with authors. Thanks for your contributions!

Best, AC

We thank the AC for handling our submission. We believe our responses during the rebuttal period have addressed the reviewers' main concerns in the following two areas:

  1. Theoretical concerns. Reviewer LRpF raised several concerns about our theoretical analysis. Following our responses, the reviewer expressed satisfaction with the clarifications and increased their score. Reviewer uD7T did not raise any further theoretical concerns after our response.

  2. Empirical robustness. Reviewer uD7T raised concerns about adversarial strategies, including query reordering and adaptive querying. We clarified why query reordering does not compromise our mechanism, and the reviewer agreed with this explanation. We also further clarified our experimental setting, which we believe addresses the concerns about adaptive querying.

We thank the AC again for their time, support, and careful consideration of our submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.