acceptodds
Under review as a conference paper at ICLR 2027

Distillation Lineage Inspector: Black-Box Auditing of Model Distillation in LLMs

Abstract

Model distillation has emerged as a widely used technique for creating efficient models tailored to specific tasks or domains. However, its reliance on knowledge from foundation models raises significant legal concerns regarding intellectual property rights. To address this issue, we propose the Distillation Lineage Inspector (DLI) framework, which enables model developers to determine whether their large language models (LLMs) have been distilled without authorization, even in black-box settings where training data and model architecture are inaccessible. DLI is effective across both open-source and closed-source LLMs. Experiments show that DLI achieves 80% accuracy with as few as 10 prompts in fully black-box settings and yields a 45% improvement in accuracy over the best baseline under standard experimental conditions. Furthermore, we analyze how auditor knowledge of target models influences performance, providing practical insight for building privacy-preserving and regulation-compliant AI systems.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model provenance, knowledge distillation, black-box auditing, large language models, intellectual property, model lineage

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

The paper proposes a novel strategy to evaluate whether a deployed model has been distilled from a target (developer’s own model) without authorization. The proposed technique combines token and model level specifications to determine a distillation case. The pipeline starts by building distilled proxies, identifying model-style tokens, encode behavioral signatures via extracted hidden features, and concludes with a trained editor to determine if the suspicious model was likely distilled from the teacher/developer model.

Critical strengths
  • The paper problem formulation is strong, especially with the black-box setting, which enforces the framework applicability.
  • The paper is also well written and flows nicely, with board evaluations and ablations.
  • Combining both model and token level specifications to learn distilled signatures for the shadow models is a smart approach.
Critical weaknesses
  • The only thing I’m worried about is the false positives, which haven’t been explored fairly throughout the paper. Since models might end up learning similar embeddings, features, and thereby output distributions, how can the authors ensure that evaluation results are clear of false positives?
  • The evaluation is a bit weak in its current format, only two datasets and four tiny models. It’s unclear how the proposed framework is robust when evaluated on complex datasets and models (e.g., overparameterized instead of a tiny model). A board evaluation would be appreciated to understand how successful and to what extent the framework is in identifying model distillation cases.
  • The choice behind the baselines to which the proposed framework is compared against is also unclear. Are the three reported baselines enough and fully covering or is the paper missing any of existing strong baselines?
  • Performances seem to collapse under the back translation defense, it’d be great if the authors can provide logical and concrete explanation of this explained behavior.
Questions for authors
  • Please provide a discussion, clarification, or even additional results to support the robustness against false positives or how well the proposed framework handles them.
  • Please expand the evaluation to at least 1 more complex and challenging dataset
  • Please explain your rationale behind the baselines choices.
  • Please provide a conclusion of the adaptive adversary results.
2 comments

We thank the reviewer for the detailed and constructive feedback. We address these concerns below.

W1 & Q1: Robustness against false positives.

We would like to emphasize that our original submission already provides extensive false-positive evaluations. As shown in Tables 3, 7, and 9--14, DLI achieves a 0% false-positive rate under KtK_t across all four teacher models and both datasets. Overall, across 80 negative audit decisions, DLI achieves FPRs of 0% under KtK_t and 2.5% under KwK_w.

To further strengthen the false-positive evaluation, we consider students distilled exclusively from alternative teachers as challenging negative samples. Specifically, each teacher-specific auditor is evaluated against 30 students distilled from the other three teachers (10 per teacher).

The following matrix reports the proportion of students classified as originating from the target teacher under KtK_t. Diagonal entries represent TPR, while off-diagonal entries represent FPR.

\text{Auditor / Student Source} & \text{LLaMA3} & \text{GPT-4o} & \text{Pythia} & \text{Qwen2.5} \\\\ \hline \text{LLaMA3} & \mathbf{1.00} & 0.00 & 0.00 & 0.00 \\\\ \text{GPT-4o} & 0.00 & \mathbf{0.80} & 0.00 & 0.10 \\\\ \text{Pythia} & 0.00 & 0.00 & \mathbf{0.90} & 0.00 \\\\ \text{Qwen2.5} & 0.00 & 0.10 & 0.00 & \mathbf{0.70} \end{array}$$ These results provide stronger evidence that DLI captures target-teacher-specific distillation signatures rather than generic similarities arising from domain adaptation or model architecture, while maintaining a low false-positive rate. We will incorporate these additional negative controls and their evaluation details into the revised manuscript. W2 \& Q2: Evaluation on more challenging datasets and models. We would like to clarify that our evaluation includes four diverse teacher models, including LLaMA3-8B, Pythia-12B, Qwen2.5-7B, and GPT-4o-mini, along with 10 distinct student architectures. Thus, our evaluation is not limited to four tiny models. Importantly, DLI identifies distillation lineage through teacher-specific token probability signatures rather than downstream task accuracy. Therefore, task difficulty does not necessarily correspond to auditing difficulty. Nevertheless, more challenging generation tasks may introduce different output distributions and linguistic patterns, making them valuable for evaluating DLI's generalizability. To further strengthen our evaluation, we additionally consider the Finance-Alpaca dataset, which comprises approximately 69,000 financial instruction-response pairs covering diverse finance-related questions and tasks. Using GPT-4o-mini as the target teacher, DLI achieves a detection accuracy of 0.95 under $K_t$ and 0.90 under $K_w$ in this illustrative setting. These results demonstrate that DLI maintains strong detection performance on challenging QA tasks, further supporting its generalizability across different generation settings. We will incorporate the additional evaluation into the revised manuscript. W3 \& Q3: Rationale for baseline selection. We would like to clarify that the rationale for our baseline selection is already discussed in Section 2 (Related Work), Section 5.1 (Experimental Setup), and Appendix A.5 (Baseline Implementation Details). Specifically, we select three representative approaches to model provenance tracing that capture distinct types of behavioral evidence: Matching Pairs[1] examines input–output embedding representations, Model Provenance Testing[2] uses statistical hypothesis testing to assess model similarity, and Part-of-Speech Templates[3] exploits syntactic patterns inherited through distillation. We also provide implementation details in Appendix A.5, including the adaptations necessary to evaluate these methods under our binary lineage-verification setting. Thus, our baseline selection is grounded in the most relevant existing provenance-tracing approaches rather than an arbitrary choice of comparison methods.

W4 & Q4: Explanation of back-translation degradation and adaptive adversary results.

We would like to clarify that this behavior is already discussed in Section 5.3 (Impact of Adaptive Adversary), with results reported in Table 4.

The degradation under back-translation follows directly from DLI's underlying mechanism. DLI relies on teacher-specific token probability signatures inherited during distillation. Back-translation modifies teacher-generated responses before distillation, preserving semantic content while altering lexical choices, sentence structures, and token-level patterns. Consequently, students trained on the transformed corpus inherit weaker teacher-specific signatures, making lineage detection more challenging.

Overall, the results demonstrate that DLI is relatively robust to partial stylistic modifications, such as cross-teacher fine-tuning, inference-time rewriting, and lightweight logit perturbation, but remains vulnerable to systematic disruption of teacher-specific token distributions before distillation.

[1] Foley, Myles, et al. "Matching pairs: Attributing fine-tuned models to their pre-trained large language models." Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics. 2023.

[2] Nikolic, Ivica, Teodora Baluta, and Prateek Saxena. "Model provenance testing for large language models." Advances in Neural Information Processing Systems 38 (2026): 34126-34153.

[3] Wadhwa, Somin, et al. "Who taught you that? Tracing teachers in model distillation." Findings of the Association for Computational Linguistics: ACL 2025. 2025.

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

The paper proposes Distillation Lineage Inspector (DLI), which tests whether a suspect LLM was distilled from a specified teacher using black-box access. DLI constructs shadow students, selects tokens from teacher-generated text that have low likelihood under a comparison model, and trains an auditor on histograms of the suspect model's probabilities for these tokens. It considers both direct token-probability access (K_t) and a text-only setting (K_w) that estimates probabilities through repeated generation. Experiments use four teachers, ten student base models, and healthcare and legal datasets, with comparisons to prior provenance methods and several evasion interventions.

Critical strengths
  1. The problem is relevant, and the access assumptions are practically motivated. Auditing distillation across different model architectures without access to suspect weights or training data is valuable. Considering a text-only interface in addition to token probabilities broadens the potential applicability.

  2. The feature construction is concrete, and the empirical study includes multiple teachers, student architectures, two domains, baseline comparisons, and useful sensitivity analyses. The strong results in the standard configurations suggest that the selected token probabilities contain detectable behavioral signals. Including back-translation and post-distillation fine-tuning is also helpful for exposing limitations beyond the standard setting.

Critical weaknesses
  1. The evidence does not yet establish reliable teacher-specific lineage detection, especially after subsequent modification.

The main evaluation compares students distilled from a target teacher with their original, non-distilled counterparts (Sections 5.1-5.2). This leaves domain adaptation and the distillation training procedure confounded with teacher identity. Evaluating four teachers separately does not show that an auditor for teacher T rejects students trained exclusively from another teacher T' on the same domain. A model fine-tuned on ordinary in-domain responses is another important negative control. Without these matched controls, the results remain compatible with detecting domain adaptation or generic teacher-generated training rather than a signature specific to T. This is an unresolved alternative explanation, not evidence that DLI necessarily learns only domain adaptation. The cross-teacher fine-tuning intervention in Table 4 does not resolve this issue: those models were first distilled from T and therefore remain positive examples of T lineage.

The existing evasion results also do not establish robust persistence of that signal. Table 4 already includes post-distillation cross-teacher fine-tuning; for the Llama teacher, accuracy falls from 1.00 to 0.75 under K_t and from 0.85 to 0.70 under K_w. Back-translation before distillation reduces these accuracies to 0.55 and 0.50, respectively, although this tests a different stage of evasion. Appendix Figure 6 further shows K_w accuracy near 0.50 with greedy or beam decoding. These are informative limitations. However, the paper does not demonstrate resistance to fine-tuning explicitly optimized against the auditor or a surrogate auditor, nor jointly quantify detection, retained task utility, and attack cost. Consequently, the current evidence neither establishes resilience to such evasion nor proves that fingerprints can always be removed cheaply without utility loss. The practical robustness claims need to be narrowed or supported more directly.

  1. The demonstrated methodological advance over prior work is limited.

Teacher inference from distilled model outputs is already studied directly by Wadhwa et al., "Who Taught You That? Tracing Teachers in Model Distillation" (Findings of ACL 2025). Foley et al., "Matching Pairs" (ACL 2023), and Nikolic et al., "Model Provenance Testing for Large Language Models" (NeurIPS 2025), also establish related black-box model-provenance approaches, although they primarily address other derivation settings. DLI cites and compares with these works, which I appreciate. My concern is the size and validation of the increment, rather than missing citations. The clearest additions are the selected-token probability features, shadow-student training, and their use in the two access regimes. Moving from closed-set teacher identification to binary verification is useful, but calling the task open-world does not itself establish reliable rejection of unknown or other-teacher students. The claim to initiate distilled-model provenance auditing is too broad given the direct precedent. Stronger teacher-specific controls are particularly important for showing that this feature-based extension provides substantial new capability.

Overall, the topic is worthwhile and the standard-setting results are promising. However, the current evaluation does not sufficiently distinguish target-teacher lineage from other-teacher distillation or ordinary domain fine-tuning, and the evasion experiments leave the practical persistence of the signal uncertain. Given the direct prior work, I also find the demonstrated methodological increment modest. These concerns affect the central contribution and currently outweigh the strengths. My recommendation could improve with convincing matched negative controls and clearer evidence about the robustness-utility trade-off.

Questions for authors
  1. Can the authors provide a small cross-teacher audit matrix using the existing teachers and student architectures? For each auditor targeting T, evaluate students trained exclusively from T' on matched prompts and training budgets, including T' != T. Add an ordinary in-domain fine-tuning control with the same student initialization and training budget. Keep suspect models separate from shadow-model training and threshold selection. Model-level false-positive rates and true-positive rates at a fixed false-positive rate would directly address whether the detector identifies T rather than generic domain adaptation. If these controls already exist, please identify the protocol and results.

  2. What robustness claim is justified by the post-distillation fine-tuning results? Please clarify the data size, training budget, and resulting task utility for the existing cross-teacher fine-tuning intervention. A bounded extension on one existing teacher-domain pair would be useful: fine-tune a student to reduce the DLI score using an accessible auditor or surrogate, then evaluate the frozen detector on fresh, held-out audit prompts. Report detection alongside retained domain utility and attack/query cost, and distinguish detector-aware adaptation from the generic interventions already tested. This would test the stated auditing setting without requiring a new benchmark suite or a universal security guarantee.

  3. What substantive capability does DLI add beyond the cited teacher-attribution and provenance methods once teacher identity, adaptation, and query budget are controlled? Please separate the contribution of token selection from shadow-student training, and qualify the claimed open-world advantage using the other-teacher negatives above. Any comparison should use matched access assumptions and account for the repeated generations used by K_w; a prompt count alone is not a complete query budget.

The first two questions concern the validity and practical scope of the current claims; the third concerns the size of the contribution. Clear teacher-specific discrimination and a credible robustness boundary could change my assessment. I am not requesting a new defense algorithm or resistance to every possible adversary.

5 comments

We thank the reviewer for the detailed and constructive feedback. We address these concerns below.

W1 & Q1: Teacher-specific detection.

We would first like to clarify that our work addresses a specific black-box auditing question: "Is the suspicious model distilled from the target LLM?" Accordingly, our primary experiments focus on distinguishing distilled students from their original base models, as the base models are arguably the most similar to the students obtained through lightweight distillation. We aim to evaluate whether DLI can reliably detect distillation lineage while maintaining a low false-positive rate, even under such challenging conditions.

We agree that including additional fine-tuned models as negative examples would strengthen the evaluation. To this end, we conducted additional experiments in which we fine-tuned models on the MedQuad dataset, which belongs to the same healthcare domain as our original dataset, using the same training configurations. The detection performance remains comparable for Qwen2.5, GPT-4o, and Pythia. Notably, for LLaMA3, the detection accuracy improves to 0.90 under KwK_w, as one previously misclassified negative sample is now correctly classified as negative.

To further evaluate teacher specificity, we consider students distilled exclusively from alternative teachers as negative samples. For each target teacher TT, its DLI auditor is evaluated against 10 students distilled from TT and 30 students distilled from the other three teachers T′T' (10 per teacher ). The following matrix reports the fraction of students classified as originating from the target teacher under token-level access (KtK_t) HealthCareMagic dataset. Diagonal entries represent TPR, while off-diagonal entries represent FPR.

\text{Auditor / Student Source} & \text{LLaMA3} & \text{GPT-4o} & \text{Pythia} & \text{Qwen2.5} \\\\ \hline \text{LLaMA3} & \mathbf{1.00} & 0.00 & 0.00 & 0.00 \\\\ \text{GPT-4o} & 0.00 & \mathbf{0.80} & 0.00 & 0.10 \\\\ \text{Pythia} & 0.00 & 0.00 & \mathbf{0.90} & 0.00 \\\\ \text{Qwen2.5} & 0.00 & 0.10 & 0.00 & \mathbf{0.70} \end{array}$$ Using independently calibrated thresholds, the four auditors achieve TPR@5\%FPR values of 100\%, 80\%, 90\%, and 90\% for LLaMA3, GPT-4o, Pythia, and Qwen2.5, respectively. These results demonstrate that DLI distinguishes target-teacher students from students distilled exclusively from alternative teachers while maintaining low false-positive rates. Together with the additional MedQuAD experiments, they strengthen the evidence that DLI captures teacher-specific lineage signals rather than merely detecting generic fine-tuning. We will incorporate these results and the corresponding evaluation details into the revised manuscript.

W1 & Q2: Robustness against post-distillation modifications.

We first clarify the robustness implications of our existing cross-teacher fine-tuning experiment. As reported in Table 4, DLI maintains detection accuracies of 0.75 under KtK_t and 0.70 under KwK_w for LLaMA3 after cross-teacher fine-tuning, indicating that teacher-specific lineage signals remain partially detectable following post-distillation adaptation.

For a controlled comparison, we use the same fine-tuning budget for cross-teacher and detector-aware fine-tuning: 5,000 QA training samples, LoRA for one epoch, a learning rate of 5×10−55\times10^{-5}. Cross-teacher fine-tuning uses responses generated by an alternative in-domain teacher, whereas detector-aware fine-tuning explicitly optimizes against a surrogate auditor while preserving downstream QA performance.

To further evaluate DLI against detector-aware adversaries, we introduce a targeted fine-tuning attack consisting of three steps:

Step 1: We use independent prompts and their corresponding model-style tokens to construct a surrogate auditor following the original DLI pipeline. Since the original histogram-based AutoGluon classifier is non-differentiable, we train a differentiable MLP surrogate using soft probability-histogram features to approximate its detection scores.

Step 2: We fine-tune the distilled students using the same training budget as cross-teacher fine-tuning. The optimization combines a QA preservation loss and a detection-evasion loss:

L=LQA+λLevasion,\mathcal{L}=\mathcal{L}{QA}+\lambda\mathcal{L}_{evasion},

where λ\lambda controls the attack strength. We consider three settings: Weak (λ=0.1\lambda=0.1), Moderate (λ=0.3\lambda=0.3), and Strong (λ=0.5\lambda=0.5).

Step 3: We evaluate the modified students using the original frozen DLI auditor on held-out prompts. We additionally measure downstream QA utility retention relative to the original distilled students.

The following table illustrates the robustness-utility trade-off for LLaMA3 on the HealthCareMagic dataset:

\text{Attack Setting} & K_t\\;\text{Acc.} & K_w\\;\text{Acc.} & \text{Utility} \\\\ \hline \text{Original} & 1.00 & 0.85 & 100\\% \\\\ \text{Cross-teacher FT} & 0.75 & 0.70 & 87.2\\% \\\\ \text{Weak }(\lambda=0.1) & 0.95 & 0.85 & 98.5\\% \\\\ \text{Moderate }(\lambda=0.3) & 0.80 & 0.70 & 86.5\\% \\\\ \text{Strong }(\lambda=0.5) & 0.65 & 0.55 & 64.9\\% \end{array}$$ The illustrative results reveal a clear robustness-utility trade-off. Under weak detector-aware attacks, DLI maintains high detection accuracy with minimal utility loss. Stronger optimization progressively weakens detection performance but incurs substantially greater utility degradation. W2 \& Q3: Methodological contributions. DLI advances existing teacher-attribution methods by introducing teacher-specific token signatures combined with shadow-student learning for black-box distillation lineage verification. To isolate the contribution of token selection, we replace model-style tokens with randomly selected tokens while keeping the probing budget, shadow models, feature encoding, and auditor training unchanged. The average results across four teachers on HealthCareMagic are: $$\begin{array}{l|cc} \text{Method} & K_t\\;\text{Acc.} & K_w\\;\text{Acc.} \\\\ \hline \text{Random Token Selection} & 0.550 & 0.525 \\\\ \text{DLI (Full)} & \mathbf{0.925} & \mathbf{0.850} \end{array}$$ Model-style token selection improves detection accuracy by 37.5 and 32.5 percentage points under $K_t$ and $K_w$, respectively, highlighting the importance of token selection. Separately, the shadow ensemble-size analysis in Section 5.3 demonstrates that increasing shadow-student diversity improves detection accuracy and stability. These results highlight the complementary contributions of teacher-specific token selection and shadow-based learning.

Thank you for your clarification. If the authors clarify how model and data separation was enforced across shadow training, threshold calibration, and cross-teacher testing, and define the attack utility metric, I would be willing to raise my score.

Reviewer_bV8WOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on black-box distillation detection. It proposes Distillation Lineage Inspector (DLI), which identifies teacher-specific tokens and trains an auditing classifier using probability features.

Critical strengths

The problem is interesting and practically relevant. It is a useful research direction that detecting distillation without access to the suspect model’s parameters or training data.

The experiments cover multiple teachers, student architectures, and two domains. The paper also considers text-only access and reports several limitations through robustness experiments.

Critical weaknesses

My main concern is whether DLI identifies the specific teacher or simply detects domain fine-tuning. The main experiments compare distilled students with their original base models. Models fine-tuned on the same domain using another teacher or human-written answers would provide stronger negative examples. Without these controls, high accuracy does not clearly establish teacher-specific detection.

The coverage of distillation pipeline is limited. The paper only covers a pipeline as LoRA fine-tuning on teacher-generated responses. The generalizability claims therefore seem broader than the evidence supports.

I am not sure whether the baseline comparison is fair enough. A fairer comparison would use the same shadow models and query budget for both proposed methods and MP method.

Questions for authors

Can DLI distinguish students of the target teacher from models trained on the same domain using another teacher or human-written answers?

Does detection still work when the shadow students and suspect models use different distillation procedures?

2 comments

We thank the reviewer for the constructive feedback and clarify the concerns below.

W1 & Q1: Teacher-specific detection vs. domain fine-tuning.

We would first like to clarify that our work addresses a specific black-box auditing question: "Is the suspicious model distilled from the target LLM?" Accordingly, our primary experiments focus on distinguishing distilled students from their original base models, as the base models are arguably the most similar to the students obtained through lightweight distillation. Our goal is to evaluate whether DLI can reliably detect distillation lineage while maintaining a low false-positive rate, even under such challenging conditions.

We agree that including additional fine-tuned models as negative examples would strengthen the evaluation. To this end, we conducted additional experiments in which we fine-tuned models on the MedQuad dataset, which belongs to the same healthcare domain as our original dataset, using the same training configurations. The detection performance remains comparable for Qwen2.5, GPT-4o, and Pythia. Notably, for LLaMA3, the detection accuracy improves to 0.90 under KwK_w, as one previously misclassified negative sample is now correctly classified as negative.

These results further support DLI's effectiveness in distinguishing target-teacher distillation from other fine-tuning settings.

W2 & Q2: Generalization across distillation pipelines.

Our original experiments employ sequence-level knowledge distillation, a widely used approach for constructing shadow and suspect models. To further evaluate the generalizability of DLI across different distillation procedures, we additionally reconstructed the shadow models using preference-based distillation via Direct Preference Optimization (DPO), while keeping the evaluation settings unchanged.

We then repeated the experiments on the Healthcare dataset. The results show that both KtK_t and KwK_w achieve performance comparable to the original distillation setting. Specifically, DLI achieves accuracies of 1.00, 0.95, 0.90, and 0.90 under KtK_t, and 0.85, 0.85, 0.90, and 0.95 under KwK_w, for LLaMA3, GPT-4o, Pythia, and Qwen2.5, respectively. These results suggest that DLI is not strongly dependent on a specific distillation procedure, demonstrating its robustness across different distillation settings.

We will include these additional results in the revised manuscript to further demonstrate the generalizability of DLI.

W3: Fairness of the baseline comparison.

We would like to clarify that both DLI and MP use exactly the same shadow models and query budget (1,000 prompts) to ensure a fair comparison. As described in Appendix A.5 (Baseline Implementation Details), we adapted MP to our open-world setting by training a binary classifier on input-output embeddings from the target teacher as positive samples and those from independent shadow models as negative samples. We will further clarify these experimental settings in the revised manuscript to avoid any confusion.

The cross-teacher test is the most useful addition for me. It makes the case for teacher-specific detection much more convincing when token probabilities are available. I still don’t see the same check for text-only access, and the low false-positive results need some caution given the small number of models and how thresholds were chosen. The DPO experiment helps, but only changes the shadow training. I’m willing to move to weak acceptance, with these limits made clear and the actual API cost better explained.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.