acceptodds
Under review as a conference paper at ICLR 2027

Cambium: Pruning-Robust Black-Box Fingerprinting for Large Language Models

Abstract

Large language models (LLMs) are powering a wide range of real-world applications. Making them valuable intellectual property requires substantial resources to train. Therefore, fingerprints are often embedded into LLMs to verify model ownership. In local deployment, LLMs are commonly fine-pruned to reduce storage cost and accelerate inference, especially in resource-constrained settings such as edge devices and small or medium-sized organizations. However, fine-pruning can erase embedded fingerprints and make model ownership unverifiable. We propose Cambium, a pruning-robust black-box fingerprinting method for LLMs. Our key observation is that a fingerprint is more likely to survive fine-pruning when its gradient is more consistent with the task gradient. To encourage this gradient consistency, Cambium first generates a verifiable and task-compatible fingerprint pair by selecting a target response supported by the model's natural output distribution. It then embeds the fingerprint through gradient-modulated supervised fine-tuning, which encourages consistency between the fingerprint gradient and the model task gradient. Comprehensive experiments with four LLMs on two different task datasets show that Cambium effectively embeds fingerprints, achieving a 100% fingerprint success rate (FSR) while preserving model utility. Across five increasingly aggressive fine-pruning settings, Cambium maintains an FSR of 71.8%–100%, while the three baselines reach 0% FSR in 46.7% of the evaluated settings. To verify ownership, Cambium achieves a 100% verification success rate (VSR) across all fine-pruning settings with at most five black-box queries.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Model Ownership Verification, Black-Box Fingerprinting

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_6AmwOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new large language model fingerprinting scheme, which can be robust to fine-pruning operations. First, they create the fingerprinting pair. Then they find the most similar input-output pair from the normal corpus and constrain the gradient similarity between the normal pair and the fingerprint pair. This makes the fingerprinting pair hard to be forgotten by model fine-pruning. Experiments show the effectiveness of the fingerprinting design.

Critical strengths
  1. This paper's logic is easy to follow, and the problem formulation is very clear.

  2. The fingerprinting design is reasonable and intuitive, simple but effective.

Critical weaknesses
  1. I have concerns about the reported 0% false positive rate. Since Cambium uses semantic embedding similarity for verification, clean models may also produce responses similar to the target fingerprint. Particularly, fingerprint gradient are aligned with some normal tasks.

  2. Section 3.1 could be clearer by explicitly defining the threat model, like the attacker's knowledge and capabilities.

Questions for authors
  1. Could clean models produce semantically similar fingerprint responses and pass the verification threshold?

  2. Please clarify the threat model in Section 3.1 to improve the paper's clarity.

2 comments

Thank you for your thoughtful review and for recognizing the clarity and rationale of Cambium. We sincerely appreciate your suggestions and the opportunity to clarify your concerns. We respond to each point below.

W1 / Q1. Can a clean model pass semantic verification?

Thank you for raising this question. We would like to clarify that although Cambium selects target tokens from the model's natural output distribution, it excludes the highest-probability tokens at each step after the first token. Repeatedly selecting these less likely tokens makes the complete target fingerprint response unlikely to be generated naturally by the clean model. Table 1 provides evidence that clean-model responses also remain far from the target responses in embedding similarity. Across four models on two task datasets, the mean embedding similarities range from 0.1273 to 0.2006, well below the default verification threshold of 0.7.

We further expanded the evaluation to test whether models without the fingerprint could pass semantic verification. For each of four source models, we selected 100 additional fingerprint pairs using Cambium's generation procedure and queried the corresponding clean model and each of the other three models 100 times per prompt. At the default threshold of 0.7, no false positives were observed across 160,000 queries. These results provide further evidence that responses from clean and unrelated models are unlikely to pass semantic verification against the target fingerprint responses under the evaluated conditions.

Queried modelsQueries across all four source modelsFalse-positive responsesObserved FPR
Corresponding clean models40,00000%
Unrelated models120,00000%

In the revision: Appendix A.3 and Table 5 report the evaluation and clarify that the observed 0% FPR applies to the evaluated pairs and models. Section 5.4 links it to the default verification settings.

W2 / Q2. Attacker knowledge and capabilities

Thank you for your suggestion to clarify the threat model. We agree that the attacker's knowledge and capabilities should be clearly specified.

Section 3.1 focuses on the central problem studied in this paper: the damage to embedded fingerprints caused by fine-pruning during model deployment. Due to space constraints, we discuss malicious attacks aimed at weakening or removing fingerprints in Appendix A.6. This appendix details the attacker's knowledge and capabilities, the specific attacks considered, and their experimental evaluation. The attacker has white-box access to the model and may know the fingerprinting algorithm, training procedure, and verification rule, but does not know the owner's secret fingerprint prompts or target responses. We evaluate continued fine-tuning, model merging, quantization, and fingerprint overwrite attacks.

In the revision: Section 3.1 clarifies the deployer's access, confidentiality of the pairs, and fingerprint embedding before fine-pruning. Appendix A.6 explicitly describes adversarial knowledge and capabilities.

Dear Reviewer 6Amw,

Thank you again for your encouraging assessment and thoughtful feedback. We greatly value your comments and would appreciate hearing whether our responses have satisfactorily addressed your concerns.

We have approached this research with great care and have given the same careful attention to the points you raised. If our responses have resolved your concerns, we would appreciate it if you could take these clarifications into account in your final assessment.

We would be happy to provide any further clarification you may need. Your feedback is very valuable to us. Thank you again for your time and consideration.

Best regards, The Authors

Reviewer_n8CFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies black-box ownership verification for LLMs that are compressed by structured pruning and then fine-tuned on task data. The authors observe that a fingerprint is more likely to survive fine-pruning when its gradient is more aligned with the task gradient, and propose Cambium on this basis. Cambium builds the fingerprint response by sampling every token after the first from probability ranks 10–20 of the clean model, embeds the pair with LoRA under a penalty that maximizes the cosine similarity between the fingerprint gradient and the gradient on retrieved task samples, and verifies ownership by sentence-embedding similarity between the returned and target responses, accepting when at least p_ver of q_ver queries pass. On four 2B–7B LLMs and two task datasets, with LLM-Pruner at five levels, Cambium keeps 71.8–100% FSR after fine-pruning, while IF, SF, and SCW reach 0% FSR in 56 of 120 settings. With p_ver = 1, it reaches 100% VSR with at most five queries.

Critical strengths
  • The threat model is concrete: a third party that prunes may have white-box access but does not know the secret pairs, and recovery fine-tuning uses only task data. Defining the five pruning levels by perplexity degradation instead of a fixed pruning ratio makes them comparable across models.
  • The gap to the baselines is large. IF, SF, and SCW fall to 0% FSR in 56 of 120 fine-pruning evaluations, while Cambium stays at or above 71.8%. The ablation separates the two components: on Qwen2.5-3B at L5, using standard SFT instead of gradient-modulated SFT lowers FSR from 96.3% to 2.5% on Dolly and from 71.8% to 14.7% on C4.
  • The appendix reports results that are less favorable to the method instead of leaving them out. FSR falls to 54.0% for Gemma-2B (Dolly) after ten epochs of continued fine-tuning and to 33.5% for Qwen2.5-3B (C4) at merging ratio 0.75. ShortGPT, two more datasets, and decoding temperature are also covered.
Critical weaknesses
  1. The 0% false positive rate behind p_ver = 1 rests on thin evidence. Sec. 5.4 sets p_ver = 1 based on the observed 0% FPR, but App. C shows that this FPR comes from two fingerprint pairs per model with 100 queries each. Zero acceptances in 200 queries bounds the per-query FPR only below about 1.5% (one-sided 95%), which with q_ver up to 5 does not exclude a false acceptance rate near 7% for the full protocol. Fig. 4 and App. A.4 and A.7 add conditions, but the negatives remain the same model, clean or fine-pruned; unrelated models are not tested, and no false acceptance rate is reported for the protocol itself. Please report it with confidence intervals, using many independent fingerprint prompts sent to the clean model and to unrelated models.

  2. Earlier evidence on pruning robustness of injected fingerprints is not discussed. Sec. 2.2 states that existing injected fingerprints remain vulnerable to fine-pruning, yet Yamabe et al. (2025), which the paper cites for merge resistance, reports that its fingerprint survives unstructured magnitude pruning up to a ratio of 0.5 (its Table 4). That is not fine-pruning, so I do not claim it would survive the paper's setting, but the current evidence cannot rule this out. Please add it as a baseline in a reduced setting (for example, one model at L1–L5) or state how the evaluated scope differs. This point alone would not change my recommendation.

  3. The side effects of embedding are only partly assessed. Table 1 relies on PPL, MMLU, and HellaSwag; the latter two are multiple-choice benchmarks that say little about generation quality, and PPL is measured on Dolly and C4, the datasets used for embedding and recovery. Please add a generation-based evaluation, such as MT-Bench or an AlpacaEval-style win rate, for at least one model before and after embedding.

  4. The checkpoint type of the models is unclear. The paper does not say whether base or instruction-tuned checkpoints are used (Sec. 5.1, App. C; App. A.4 mentions the Qwen2.5-3B-Instruct generation config only as an example). If they are base models, the results may not transfer to the instruction-tuned models that are often deployed. Please state the checkpoints and, if they are base models, reproduce a small part of Table 2 (for example, L1–L5 on Dolly) on one instruction-tuned model.

Questions for authors
  1. Which checkpoints (base or instruction-tuned) are used? Table 1 defines "Clean" as the model before embedding, yet its MMLU differs between the Dolly and C4 rows for all four models (e.g., Gemma-2B 0.40 vs. 0.42) and its HellaSwag for three (e.g., Mistral-7B 0.79 vs. 0.81). Is the clean model fine-tuned on each task dataset, or do the differences come from the evaluation (N_eval is not given)?
  2. Sec. 5.2 and App. A.2 say that the clean-model embedding similarity is "at most 0.2006" and "only 0.2006". Is this a mean or a maximum over queries, pairs, and seeds? Fig. 4 shows that the worst-case clean-model FPR over Dolly, C4, and L1–L5 is non-zero until τ_s reaches 0.5–0.6.
  3. For Fig. 1, on which parameters are the gradients computed, how many prompts and candidate responses are used per setting, and what is the variance?
  4. Which values of α_fp, β_pena, the LoRA rank, the learning rate, and the number of epochs are used? I could not find them. Does Cambium-SFT in Table 3 keep the L_task term of Eq. (7)?
3 comments

Thank you for your careful and detailed review. Your thoughtful comments and constructive suggestions are very valuable to us. We have carefully considered each concern and question and explain the corresponding revisions in our responses below.

W1. False positives on clean and unrelated models

We expanded the negative-model evaluation to test whether the observed zero FPR extends beyond the original fingerprint pairs and clean models (Appendix A.3, Table 5). For each of four source models, we selected 100 additional fingerprint pairs constructed using Cambium's generation procedure and queried each of the four evaluated models 100 times per prompt. No false positives occurred across 160,000 queries at the default similarity threshold of 0.7.

No false positives were observed across 160,000 queries using the additional fingerprint pairs on clean and unrelated models, providing further evidence that the default verification threshold can distinguish fingerprinted models from models without the fingerprint.

Queried modelsQueries for each source modelQueries across all four source modelsFalse-positive responsesObserved FPR
Corresponding clean model10,00040,00000%
Three unrelated models30,000120,00000%
Total40,000160,00000%

The additional evaluation also allows us to assess false acceptance under the five-query verification rule. For each of the 100 additional fingerprint prompts, we consider the first five responses and accept ownership if at least one response passes verification, following the rule used in Section 5.4. Since none of the responses passed verification, there were no false acceptances in these 100 verification trials on either the corresponding clean model or any of the three unrelated models. Assuming independent trials across fingerprint prompts, the one-sided 95% exact binomial upper confidence bound is 2.95% for each tested model. We calculate this bound separately for each source model and each queried model, rather than treating repeated queries for the same prompt as independent trials.

In the revision: Appendix A.3 and Table 5 describe the evaluation using additional fingerprint pairs. Appendix A.3 also reports the false acceptance results under the five-query verification rule and the corresponding confidence bound. Section 5.4 refers to this evidence when explaining the verification setting.

W2. Prior pruning evidence and the MergePrint baseline

Following your suggestion, we added a comparison with MergePrint on Qwen2.5-3B across L1–L5 on both Dolly and C4 (Appendix A.10). At L5, MergePrint retains 2.5% and 0% FSR, respectively, compared with 96.3% and 71.8% for Cambium. This directly evaluates MergePrint within our fine-pruning setting.

MethodDolly, L5 FSRC4, L5 FSR
MergePrint2.5%0%
Cambium96.3%71.8%

In the revision: Appendix A.10 discusses the distinction and Table 9 reports the complete L1–L5 results. Cambium's corresponding results are in Table 2.

W3. Generation quality before and after embedding

We added MT-Bench evaluation before and after fingerprint embedding across four models on two task datasets (Table 1 and Section 5.2). The largest relative score decrease is 3.6%.

ModelDolly: Clean → FPC4: Clean → FP
Gemma-2B4.96 → 4.845.33 → 5.31
Qwen2.5-3B6.07 → 5.996.15 → 5.98
Llama-2-7B5.74 → 5.684.99 → 4.81
Mistral-7B6.13 → 6.225.94 → 6.11

In the revision: Table 1 adds MT-Bench, Section 5.2 summarizes the utility changes, and Appendix C describes the evaluation.

W4. Instruction-tuned checkpoint

We extended the L1–L5 evaluation to Qwen2.5-3B-Instruct on both Dolly and C4 (Appendix A.9). Cambium maintains 100% FSR at L1–L2 and retains 68.3% and 64.3% FSR, respectively, at L5.

DatasetOriginalL1L2L3L4L5
Dolly100%100%100%100%83.5%68.3%
C4100%100%100%97.8%79.2%64.3%

In the revision: Section 5.1 clarifies the base checkpoints; Appendix A.9 and Table 8 report the instruction-tuned results.

Q1. Definition of Clean and benchmark sample sizes

The main experiments start from base checkpoints, which are first fine-tuned on Dolly or C4. “Clean” denotes the resulting task-adapted model before fingerprint embedding. Thus, the Dolly and C4 rows correspond to different task-adapted models, explaining their different MMLU and HellaSwag scores. We have clarified this in Section 5.1, the Table 1 caption, and Appendix C. We also specify the evaluation sample sizes and prompting settings: MMLU uses 1,425 fixed test questions, comprising 25 questions per subject across 57 subjects, with five-shot demonstrations from the corresponding development splits. HellaSwag uses 2,000 fixed validation questions in a zero-shot setting.

BenchmarkEvaluation samplePrompting
MMLU1,425 fixed test questions: 25 from each of 57 subjectsFive-shot, with demonstrations from the subject's development split
HellaSwag2,000 fixed validation questionsZero-shot; length-normalized likelihood

In the revision: Section 5.1 and the Table 1 caption clarify Clean; Appendix C specifies these evaluation settings.

Q2. Meaning of the clean-model similarity value 0.2006

For each clean model on each task dataset, the reported embedding similarity is averaged over query responses. The value 0.2006 is the highest of these mean embedding similarities reported in Table 1.

In the revision: Section 5.2, Appendix A.2, and the Table 4 caption clarify that these values are mean embedding similarities.

Q3. Figure 1: gradient parameters, candidate pairs, and variability

For each model on each task dataset, we generated 100 candidate fingerprint pairs, each containing a distinct prompt and one randomly generated response. We selected the pairs with the highest and lowest gradient cosine similarities, embedded them separately, and evaluated them under L1–L5. Both gradients are computed with respect to the trainable LoRA parameters in the query and value projections across all Transformer layers. We have added these experimental details to Section 4.1. Regarding variance, we ran the experiment once for each model on each task dataset, so we do not have an estimate of variability across repeated runs. The 100 candidates form the fingerprint pair selection pool, rather than repeated evaluations. Figure 1 shows a consistent trend across four models on two task datasets, supporting the observation that motivates the method.

In the revision: Section 4.1 specifies the candidate fingerprint pairs, selection procedure, and trainable LoRA parameters used to compute gradients.

Q4. Hyperparameters and the Cambium-SFT objective

Appendix C now specifies the fingerprint-loss weight of 1.0, gradient-consistency penalty weight of 0.1, LoRA rank of 16, learning rate of 10⁻⁴, and training duration of 10 epochs. Cambium-SFT retains both the task loss and the fingerprint loss, removing only the gradient-consistency penalty.

Embedding parameterValue
Fingerprint-loss weight, αfp\alpha_{\mathrm{fp}}1.0
Gradient-consistency penalty weight, βpena\beta_{\mathrm{pena}}0.1
LoRA rank16
Learning rate10−410^{-4}
Training epochs10

In the revision: Appendix C lists these hyperparameters, and Appendix A.1 explicitly gives the Cambium-SFT objective.

Thank you for clarifying. Most of my main concerns are resolved. I changed my score to 3.

Reviewer_B4peOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper observes that for watermarked models (the authors rather word it as injected fingerprints but I prefer to use this term) the watermarks are easily erased by model fine-pruning, a model compression technique based on the combination of pruning and fine-tuning. To circumvent the issue, the authors leverage a neat observation: watermarks are more robust if they are aligned with model task gradients. By aligning watermarking gradient to task gradient, they leverage this observation to produce fine-pruning proof watermarked model, and validate their approach through some experiments.

Critical strengths
  • neat observation, well presented, that provides an intuitive and good rationale to the approach
  • it works: extensive experimental coverage, with good results (up to "too good" see weakness): the method surpasses SoTA by a large margin in this fine-pruning section.
  • generally clearly written and didactic, despite some points I did not get.
Critical weaknesses
  • Unclear system model: I have some trouble understanding the general system model and its corresponding rationale. During fine-pruning, is the fine-tuning task the same than the one exploited during watermarking ? Somehow, if the owner model is doing the fine-pruning, why not simply first fine-prune, then watermark ? This would avoid all the fuss. Otherwise, if another entity (perhaps even the attacker) is doing the fine-pruning, how come the subsequent fine-tuning task is known beforehand to the owner at the time of watermarking ? And in this case, what happens if the fine-pruning tuning task is different than the one the watermark is aligned with ? Perhaps it is just a misunderstanding on my side, otherwise I believe it has deep implications on the relevance of the paper.
  • Impact is hard to assess: the importance of fine-pruning should be better emphasised. Is this a technique model owners often do, or an obscure use case rarely relevant in practice ?
  • Comparison to SoTA is only presented on fine-pruning experiments. Other attacks resistance do not include SoTA; it is then difficult to gauge the improvement over SoTA on fine-pruning is not symmetrically spent on other attacks.
  • (less critical) 100% FSR makes me suspicious. A method performance is only the product of some configuration. Reading 100% success only informs that the selected configurations were too simple, and flat 100% lines in fig 3 bring very little information overall. Regardless of the intrinsic quality of the method, I'd prefer to also learn about its limits in the paper.
  • (less critical) I was not very convinced by the limits of intrinsic fingerprints: at the end of the day, what matters is to exhibit a certificate of ownership (is this the owned model), I believe the "form" of this certificates does not matter most (we don't need an NFT gif). Moreover, TRAP for instance provides some control to the owner, letting her select the target output of the model -- which seems arguably stronger to me for a certificate that letting the owner select the target input. Including one of these methods into comparison baselines could be interesting.

minor

  • please specify temperature used for the main body experiments
  • PPL is never defined in the body, imho it would be better for self-containment.
Questions for authors

(see questions above)

  • please clarify system model
  • please clarify impact
5 comments

Thank you for your careful review and constructive suggestions, and for recognizing the observation motivating Cambium, its effectiveness under fine-pruning, and the clarity of the presentation. We greatly appreciate your comments and have carefully considered them in revising the manuscript. Our point-by-point responses are provided below.

W1 / Q1. System model and the order of fingerprinting and fine-pruning

We clarify that our work studies structured pruning followed by recovery fine-tuning to restore the model’s original task performance. Recovery fine-tuning is part of established structured-pruning pipelines, such as LLM-Pruner, not an extra requirement introduced by Cambium. Structured pruning removes model components and disrupts learned representation pathways and parameter interactions. Recovery fine-tuning is therefore a necessary step in this pipeline to restore effective interactions among the remaining components and recover task performance. In our setting, recovery fine-tuning uses the same task dataset used for fingerprint alignment. The owner therefore needs only data from the model’s original task, without anticipating an unknown future task.

Regarding a different fine-tuning task, recovery from structured pruning and adaptation to a new task address distinct problems. To examine the need for recovery, we conducted an experiment in which we skipped recovery on the original task and directly fine-tuned the pruned model on a new task. We found that the model struggled to learn the new task. This observation supports completing the recovery stage even when subsequent task transfer is intended. We therefore clarify that our focus is the complete fine-pruning pipeline: structured pruning followed by recovery fine-tuning to restore interactions among the remaining components. In the pipeline we study, subsequent adaptation to a downstream task does not replace this recovery stage.

Regarding the order of fine-pruning and fingerprinting, fine-pruning first is possible when the owner controls the entire pipeline. However, it does not address further compression by a third-party deployer after release. The deployer cannot re-embed the owner’s fingerprint without the secret fingerprint pairs, yet disclosing these pairs could enable manipulation of verification or deliberate fingerprint removal. We have clarified in Section 3.1 that the deployer has white-box access to the model and may modify its architecture and parameters, but does not know the owner’s fingerprint pairs. Cambium addresses precisely this problem: preserving an existing fingerprint through fine-pruning without requiring the deployer to know the owner’s secret pairs.

In the revision: Section 3.1 explicitly places fingerprint embedding before fine-pruning and clarifies the deployer's access and the secrecy of the pairs. Section 5.1 and Appendix C describe the task data used during recovery.

W2 / Q2. Practical relevance of structured pruning

Structured pruning is widely studied in academic research and is also used in industry to develop and deploy smaller language models. Major model developers have already released compact models built using this technique. For example, Meta developed Llama 3.2 1B and 3B by structurally pruning Llama 3.1 8B, while NVIDIA released Minitron models by pruning and distilling Llama 3.1 8B and Mistral NeMo 12B into 4B and 8B models, respectively. NVIDIA also provides structured-pruning implementations and end-to-end deployment tutorials through Model Optimizer. We have added these examples to the Introduction to make the practical relevance of structured pruning and the motivation for studying fingerprint robustness under fine-pruning clearer.

ExampleRelevance
Meta Llama 3.2 1B and 3BCompact models developed through structured pruning of Llama 3.1 8B
NVIDIA MinitronPruning and distillation of Llama 3.1 8B and Mistral NeMo 12B into 4B and 8B models; the former also illustrates compression by another organization
NVIDIA Model OptimizerStructured-pruning implementations and deployment tutorials supporting practical use

In the revision: Section 1 adds these examples and their supporting references.

W3. Baseline comparisons under other attacks

We agree that comparisons under other attacks are needed to assess whether Cambium’s fine-pruning robustness comes at the expense of resistance to other modifications. We have therefore added comparisons with IF, SF, SCW, and TRAP under general attacks in Appendix A.6. Overall, Cambium’s improved robustness to fine-pruning does not generally come at the expense of resistance to fine-tuning, model merging, or quantization attacks.

Under continued fine-tuning (Appendix A.6, Figure 6), Cambium generally performs better: after ten epochs, it matches or exceeds all four baselines across four models on two task datasets, except for Gemma-2B on Dolly, where TRAP performs better.

Under model merging (Appendix A.6, Figure 7), no method consistently dominates across the evaluated settings. Cambium retains 100% FSR on Llama-2-7B and Mistral-7B at every tested merging ratio, while IF performs better on Gemma-2B and Qwen2.5-3B when the fingerprinted model contributes only 25% of the merged weights. Thus, Cambium remains competitive, but its advantage under fine-pruning does not extend to every merging setting.

Under quantization (Appendix A.6, Figure 8), all five methods retain 100% FSR under dynamic INT8 quantization. Under BNB LLM.int8 quantization, Cambium, IF, SF, and SCW retain 100% FSR, while TRAP achieves a mean FSR of 83% across four models on two task datasets.

In the revision: Appendix A.6 adds comparisons with IF, SF, SCW, and TRAP under continued fine-tuning (Figure 6), model merging (Figure 7), and quantization (Figure 8).

W4. Interpretation of perfect scores and limitations

Reliable ownership verification after fine-pruning is the central goal of Cambium. Achieving an observed 100% verification success rate within five queries demonstrates that the verification rule meets this goal in the evaluated trials; it does not, by itself, indicate that the settings are too easy. Under the same fine-pruning settings, the four baselines reach zero FSR in 71 of 160 evaluations.

It is also important to distinguish verification success from individual-query success. Cambium’s FSR falls to 71.8% under the strongest main fine-pruning setting. The verification rule uses repeated queries precisely to accommodate this degradation, yielding the observed 100% VSR within five queries.

The original submission also reports limitations under other attacks: FSR drops to 54.0% under continued fine-tuning and 33.5% under model merging. These results in Appendix A.6 show that Cambium’s robustness is not unlimited.

W5. Intrinsic fingerprints and TRAP

We clarify that our main concern is the limited control over intrinsic fingerprints’ robustness to fine-pruning. Their preservation through fine-pruning is not explicitly optimized during fingerprint-embedding training. Nevertheless, we appreciate the reviewer’s suggestion to test this empirically and have added TRAP as a baseline in Table 2. TRAP achieves 100% FSR before fine-pruning but drops to zero in 15 of 40 fine-pruning settings, including all L5 settings. These results support our concern about intrinsic fingerprint preservation under fine-pruning in the evaluated settings.

TRAP fingerprint success rate (FSR), reproduced from Table 2. Original denotes the model before fine-pruning.

ModelDatasetOriginalL1L2L3L4L5
Gemma-2BDolly100%100%88.5%4.8%0%0%
Gemma-2BC4100%98.7%91.2%27.3%2.5%0%
Qwen2.5-3BDolly100%75.5%2.8%4.3%0%0%
Qwen2.5-3BC4100%49.2%7.7%1.5%0%0%
Llama-2-7BDolly100%73.8%54.7%0.8%0%0%
Llama-2-7BC4100%85.2%68.3%0%0%0%
Mistral-7BDolly100%100%21.2%18.8%0%0%
Mistral-7BC4100%100%35.7%15.0%6.5%0%

In the revision: Section 2.2 clarifies the distinction; Table 2 adds TRAP; Section 5.3 discusses its results.

W6. Default verification temperature

The original submission already provided the default decoding temperature in Appendix C. These details were placed in the appendix because of the main-text page limit, allowing the main text to prioritize the method and experimental results. Following your suggestion, we now also state the default verification temperature of 1 in Section 5.2.

W7. Definition of PPL

The original submission already provided definitions of PPL and the other evaluation benchmarks in Appendix C. Following your suggestion, we now also define PPL as perplexity in Section 5.1.

Dear Reviewer B4pe,

As the discussion period comes to an end, we would appreciate hearing whether our responses have satisfactorily addressed your concerns. If any points would benefit from further clarification, we would be happy to elaborate. Your feedback is very valuable to us.

Thank you again for the time and care you have devoted to reviewing our work.

Best regards, The Authors

Comments on the paper

1 comment

We thank all reviewers for their careful reading, constructive comments, and helpful suggestions. We have addressed all reviewer comments through revisions to the manuscript and the point-by-point responses below. Changes in the revised manuscript are highlighted in blue. The revisions are summarized below.

Reviewer commentsRevision and responseLocation in the revised manuscript
B4pe W1/Q1; 6Amw W2/Q2: System and threat modelClarified embedding before third-party compression, white-box access for the deployer, black-box access for the verifier, and secrecy of the fingerprint pairs; explained the recovery task and why fine-pruning before embedding does not address subsequent third-party compressionSection 3.1; Section 5.1; Appendix A.6; Appendix C
B4pe W2/Q2: Practical relevanceAdded examples of released models from Meta and NVIDIA, together with Model Optimizer support for structured pruningSection 1
B4pe W3: Baseline comparisons under other attacksAdded IF, SF, SCW, and TRAP comparisons under continued fine-tuning, model merging, and quantization attacksAppendix A.6, Figures 6–8
B4pe W4: Perfect scores and limitationsDistinguished query-level FSR from trial-level VSR; reported 71 baseline results with 0% FSR out of 160 evaluations and discussed performance degradation under other attacksSections 5.3–5.4; Table 2; Figure 3; Appendix A.6
B4pe W5: Intrinsic fingerprints and TRAPRefocused the discussion on preservation through fine-pruning and added TRAP, which reaches zero FSR in 15 of 40 settings, including all L5 settingsSection 2.2; Section 5.3; Table 2
B4pe W6–W7: Temperature and PPLStated the default verification temperature of 1 and defined perplexity in the main textSections 5.1–5.2; Appendix C
n8CF W1; 6Amw W1/Q1: False positivesEvaluated 100 additional fingerprint pairs for each source model on clean and unrelated models; observed 0% FPR across 160,000 queries; clarified that this result applies to the evaluated fingerprint pairs and modelsSection 5.4; Appendix A.3, Table 5
n8CF W2: MergePrint and prior pruning evidenceDistinguished unstructured magnitude pruning from fine-pruning and added MergePrint on Qwen2.5-3B under L1–L5 on Dolly and C4Appendix A.10, Table 9
n8CF W3: Generation qualityAdded MT-Bench before and after embedding across four models on two task datasets; the largest relative score decrease is 3.6%Section 5.2; Table 1; Appendix C
n8CF W4/Q1: Checkpoint type and instruction-tuned modelsSpecified base checkpoints for the main experiments and added Qwen2.5-3B-Instruct results across L1–L5 on Dolly and C4Section 5.1; Appendix A.9, Table 8
n8CF Q1: Clean-model definition and evaluation samplesDefined Clean as the task model obtained by fine-tuning a base checkpoint on the corresponding task dataset, before fingerprint embedding; specified the MMLU and HellaSwag sample sizes and prompting settingsSection 5.1; Table 1 caption; Appendix C
n8CF Q2: Meaning of 0.2006Clarified that 0.2006 is the largest reported mean across four models on two task datasets, not the largest embedding similarity of an individual responseSection 5.2; Appendix A.2; Table 4 caption
n8CF Q3: Figure 1 setup and variabilitySpecified the 100 candidate fingerprint pairs, selection by the highest and lowest gradient cosine similarities, and trainable LoRA parameters used to compute gradients; explained the single-run observation and lack of a repeated-run variance estimate in the response belowSection 4.1; Figure 1
n8CF Q4: Hyperparameters and Cambium-SFTAdded embedding hyperparameters and specified that Cambium-SFT retains the task and fingerprint losses, removing only the gradient-consistency penaltyAppendix A.1; Appendix C

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.