acceptodds
Withdrawn from ICLR 2027

Towards Distillation-Resistant Large Language Models: A Conditional Mutual Information Perspective

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Proprietary large language models (LLMs) embody substantial economic value and are generally exposed only as black-box APIs. However, adversaries can still exploit their outputs to extract knowledge via distillation. Previous studies have explored defense strategies, yet they primarily focus on text-based distillation, leaving the important logit-based counterpart largely unexplored. In this work, we analyze this problem from an information-theoretic perspective and present an effective solution. Specifically, we characterize distillation-relevant information in teacher outputs via the conditional mutual information (CMI) between teacher logits and input queries, conditioned on ground-truth labels. This quantity captures contextual signals beyond label supervision that can be exploited for knowledge extraction, motivating us to defend against unauthorized distillation via CMI minimization. To this end, we propose a learnable transformation matrix that purifies the original outputs and provide a theoretical analysis that establishes its effectiveness. We further derive a CMI-inspired anti-distillation objective to optimize this transformation, which effectively reduces distillation-relevant information while preserving output utility. Extensive experiments across multiple LLMs and strong distillation algorithms demonstrate that the proposed method significantly degrades distillation performance while preserving task accuracy, thereby effectively protecting models' intellectual property.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Anti-distillation, Post-processing Strategy

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_ww5FOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a defense against logit-based knowledge distillation by applying a learnable transformation to teacher logits. Motivated by conditional mutual information (CMI), the method combines a utility-preserving cross-entropy objective with a gradient-direction objective intended to disrupt student learning.

Critical strengths
  • The parameter-efficient transformation keeps the teacher weights frozen and provides a simple mechanism for modifying its outputs.
  • Experiments cover multiple student sizes and distillation objectives, with ablations and comparisons against simple logit-processing defenses.
Critical weaknesses
  • Outdated and unrepresentative evaluation. Experiments rely on Qwen2.5 and Llama 3 models and the older GSM8K, MMLU, and MATH-500 benchmarks. In particular, the standard Qwen2.5 teacher does not represent contemporary models specifically trained for extended reasoning. With open reasoning-capable models such as Qwen3.5-27B available, the absence of modern reasoning teachers and more challenging reasoning benchmarks is a major limitation. The results do not establish whether the defense remains effective when distilling stronger teachers and richer reasoning traces.
  • Incomplete threat model and missing adaptive attacks. An attacker could discard the modified logits and instead distill the teacher’s generated text, which the method explicitly aims to preserve in quality. This straightforward alternative is not evaluated, and the SFT-only baseline does not test distillation from the defended teacher’s generated responses. Furthermore, the practical relevance of assuming access to full-vocabulary logits is insufficiently justified. Consequently, the experiments support a narrower claim about disrupting selected logit-based objectives, rather than broadly preventing knowledge extraction.
  • The information-theoretic justification does not establish information removal. Theorem 2 guarantees only that deterministic post-processing cannot increase CMI. The implemented transformation (M=I+AB)(M=I+AB) is not constrained to be non-invertible; whenever it is invertible, (I(X;MZ∣Y)=I(X;Z∣Y)).(I(X;MZ\mid Y)=I(X;Z\mid Y)). Moreover, changing student gradient directions does not establish a reduction in CMI. The observed degradation could therefore reflect interference with particular optimization objectives without demonstrating the claimed removal of distillation-relevant information.
Questions for authors
  1. Does the defense remain effective against sequence-level distillation using responses generated by the defended teacher, and against attackers that retune their distillation procedure for the modified outputs?
  2. Can the authors evaluate modern reasoning teachers, such as Qwen3.5-27B, on more challenging reasoning benchmarks with appropriate generation budgets, reporting both student degradation and preservation of teacher utility?
Reviewer_vUqGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a information-theoretic defense mechanism against logit-based knowledge distillation attacks on LLMs. To prevent adversaries from extracting a proprietary LLM's core capabilities via exposed API logits, the authors introduce a transformation matrix that post-processes the teacher model's outputs while keeping its parameters frozen. By minimizing the CMI between the input queries and the logits, the proposed method effectively disrupts the student model's training. Extensive experiments demonstrate that this approach significantly degrades the reasoning and generation performance of unauthorized student models while successfully preserving the original teacher model's accuracy and text quality.

Critical strengths
  • The insights presented in the manuscript are interesting, and the proposed method is technically sound.
  • The paper is well-supported by both theoretical proofs and empirical evaluations.
Critical weaknesses
  • The manuscript lacks essential background on why API providers offer logit access.
  • While the manuscript demonstrates that the proposed method preserves the teacher model's task performance, it completely overlooks the potential negative impact on benign users who require access to logits. This omission is unacceptable and raises concerns about the practical viability of the proposed defense paradigm.
Questions for authors
  • Please survey current API providers offering logit access. Clarify their primary intent for providing logits, the supported downstream tasks, and the benign users who benefit.
  • The proposed method distorts the original logit distribution, inevitably impacting users relying on accurate logits. Could the authors evaluate this impact on logit-dependent downstream tasks? Please explicitly define the defense's applicable scenarios and practical viability if adverse effects exist.
  • Does the threat of extraction amplify or diminish as model scale increases beyond the evaluated Qwen 2.5 and Llama 3 series? Please discuss the real-world extraction risks for commercial products and the relationship between extraction risk and target model scale.
  • Following up on the previous question, can the proposed defense be seamlessly applied in real-world deployments? What potential challenges might arise, and what issues need to be resolved when scaling this defense to much larger models?
Reviewer_1jVaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Targeting on the logit-based KD, this paper propose a novel defense exploiting CMI to reduce the contextual information contained in the model output logit. Extensive experiments across multiple LLMs and strong distillation algorithms demonstrate that the proposed method significantly degrades distillation performance while preserving task accuracy, thereby effectively protecting models’ intellectual property

Critical strengths
  • The paper studies an important and timely problem.
  • Extensive evaluation covering varying models and detection methods.
Critical weaknesses
  • My main concern is that the effectiveness of the proxy signal of I(X;Z). It seems that this proxy signal depends on a paticular student model, and the Z’ are actually optimized to reduce the effective information of this model. However, different student models may have different architectures and learning dynamics, so whether Z’ can also suppress the information that other student models can utilize remains to be unclear.
  • The paper has several inconsistencies between formulas and textual descriptions, e.g., the cross-entropy in Equation 11 are inconsistent with the manuscript.
Questions for authors

Overall, this paper addresses an important issue and proposes a novel defense method for knowledge distillation based on logit. However, there is a lack of sufficient justification for the generalization of the I(X,Z) proxy: reducing the information available to a specific proxy student does not necessarily prevent other students from exploiting this information.

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

Related papers

Discussion (0)

Sign in to comment.