acceptodds
Under review as a conference paper at ICLR 2027

RAFP: Identifying LLM Lineages via Rare-Region Fingerprints

Abstract

Large language models (LLMs) are increasingly adapted and redistributed through finetuning, parameter-efficient tuning, and quantization, making it difficult to verify whether a suspect model is derived from a particular pretrained model. Existing fingerprinting approaches often rely on modifying the protected model or construct behavioral signatures without explaining why they should persist under downstream adaptation. We introduce RAFP, a non-invasive framework for identifying LLM lineages through rare-region fingerprints. Our key insight is that downstream adaptation is concentrated on common high-density language behaviors, while low-probability prompt regions receive weak optimization signal and exhibit limited gradient alignment under finetuned distribution. We show theoretically that, under mild assumptions on downstream coverage and gradient alignment, the likelihood drift of behaviors in these rare regions remains bounded during finetuning. Building on this insight, RAFP searches for rare prompt–response pairs whose behaviors are shared across models in the same lineage but remain discriminative from unrelated models, without modifying model parameters. The resulting fingerprints support black-box ownership verification and can be generated post hoc for pre-trained models. Experiments across four LLM families and multiple downstream adaptations, including supervised finetuning, LoRA, quantization, prompt-template variation, and decoding changes, show that RAFP achieves strong fingerprint persistence and substantially outperforms prior fingerprinting baselines in black-box settings.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Model Fingerprinting, Model Lineage, Model Provenance, Black-box Verification, Robustness to Downstream Tasks

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_A8e2OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a fingeprinting method to identify LLMs being served behind an API. The proposed method is non-invasive, and aims to produce unique strings which trigger a particular response for an LLM and its derivative LLMs. These unique strings are from "low probabilty" prompt regions, which are unlikely to be contained in fine-tuning datasets. As a result, the behaviour of LLMs on these fingerprints is fairly stable even after fine-tuning. These fingerprints are constructed using discrete optimization techniques like GCG, with an optional perplexity regularizer. The technique is evaluated on Llama models, and the fine-tuning persistence as well as robustness to prompt templates, decoding time changes and quantization are also reported.

Critical strengths
  1. The theoretical results saying that behaviour on rare prompts would not be disrupted under standard fine-tuning is nice. The gradient analysis is non-trivial in my view and adds to the paper. However, I am unsure if the theory connects much to practice (see below).
  2. The evaluations seem thorough, though the adversarial evaluations (esp for the robustness are lacking, see below).
  3. The paper is very well written and easy to follow.
Critical weaknesses

My main concern is with the claimed robustness of the presented scheme

  1. Fig 8 seems to imply that the scheme has limited robustness, since ab adversary could easily filter out RAFP fingerprints which look like outliers on the pplx length curves? I am unsure if this takes into account the regularizers to generate more realistic fingerprints.
  2. It would be nice if the paper could clarify why RAFP fingerprint prompts have ρ(τ)<<1\rho(\tau) << 1 (a). I believe there is some kind of an impossiblity here-- if (a) is true, an adversary can easily filter out prompts which look unnatural based on perplexity (e.g. the attack on TRAP from Nasery et al 2026). To alleviate this, the present paper incorporates a perplexity loss in section 6. However, this means that the resulting fingerprints are then closer to fine-tuning distributions that an adversary might pick, which means that fine-tuning might lead to forgetting. It would strengthen the paper to characterize the tradeoffs introduced by the perplexity loss.
  3. I also do not believe that ρ(τ)<<1\rho(\tau) << 1 is a good assumption for adversarial fine-tuning. Indeed, Xu et al 2025 (MEraser) explicitly construct fine-tuning distributions which seem to overlap with certain fingeprinting distributions. In my view it should be fairly straightforward for an adaptive adversary to construct query distributions for fine-tuning to disrupt RAFP fingerprints.
  4. Minor issue --- Missing prior work -- Gubri et al (TRAP)
Questions for authors
  1. Were the results in Fig 8 generated with the pplx loss? What are the tradeoffs in robustness that this loss brings?
4 comments
AuthorsOpenReview ↗

Addressing PPL-Filtering Vulnerability and Adaptive Fingerprint Erasure with New Robustness Evaluations

W1, W2, Q1: Perplexity filtering and tradeoff of fingerprint persistence

We thank the reviewer for highlighting the tradeoff between filtering resistance and fingerprint persistence.

We clarify that Figure 8 was generated using RAFP fingerprints without PPL regularization (λ\lambda=0). The figure illustrates that unregularized rare-region fingerprints can be distinguished from natural prompts using perplexity, motivating our ablation PPL-aware optimization shown in Section 6.

We clarify that low probability under the downstream fine-tuning distribution does not necessarily imply rejection by a practical PPL filter. Our PPL-aware optimization aims to balance these competing objectives rather than eliminate the tradeoff entirely.

To further investigate this tradeoff, we extend our evaluation to conventional GCG-QA fingerprints, with and without PPL regularization, using the same filtering conditions. As shown in our additional results (Table 1), we report both matching accuracy among fingerprints surviving the filter on both base model and SFT model.

Importantly, we discover this tradeoff is not unique to RAFP only, since conventional GCG-based methods may also produce high-perplexity trigger strings. Our additional comparison evaluates whether RAFP-PPL (λ\lambda=0.3) offers a more favorable balance between filter acceptance and persistence than GCG-QA under equivalent PPL constraints.

Table 1: Fingerprint robustness under perplexity filtering (%).

MethodPPL Rejection ↓\downarrowBase (Filtered) ↑\uparrowSFT (Filtered) ↑\uparrow
Rand. triggers90 (9/10)0 (0/1)0 (0/1)
GCG-QA w/o PPL80 (8/10)100 (2/2)50 (1/2)
GCG-QA w/ PPL20 (2/10)87.5 (7/8)50 (4/8)
RAFP w/o PPL100 (10/10)——
RAFP w/ PPL (λ\lambda=0.1)20 (2/10)100 (8/8)87.5 (7/8)
RAFP w/ PPL (λ\lambda=0.3)10 (1/10)100 (9/9)66.7 (6/9)
RAFP w/ PPL (λ\lambda=0.5)10 (1/10)88.9 (8/9)66.7 (6/9)

- Values are percentages (counts). PPL rejection is rejected/total; Base/SFT matching is successful/surviving total.

We will clarify that RAFP does not guarantee resistance to arbitrary input filtering; rather, RAFP seeks persistent fingerprint behaviors under practical filtering constraints.

W3: Robustness against adaptive fine-tuning.

We thank the reviewer for pointing out the paper Xu et al. 2025 (MEraser). We agree that an adaptive adversary could intentionally manipulate the fine-tuning distribution to increase exposure to rare-region prompts, further violating our assumption 1. However, even if an adversary knows RAFP's construction strategy, it is still hard for he/she to guess or access the full secret fingerprint pairs. In our appendix A.8, Unforgeable property, we show the probability that the adversary model erases all fingerprint prompts is very small. Even to achieve a 1% probability of erasing the fingerprints, the adversary would need to generate N = 2.71e105 query-response pairs.

Following the reviewer's suggestion, we evaluate this threat under a realistic downstream adaptive finetuning setting, we use the full ShareGPT fine-tuning dataset (50k samples) and inject rare-region samples at ratios of 1%, 2%, 5%, and 10%. These samples follow RAFP's rare-token initialization strategy but do not contain the secret verification fingerprints. In Table 2, we compare GCG-QA, RAFP, and RAFP-PPL on the resulting fine-tuned checkpoints. This experiment shows that injecting a small fraction of rare-region samples only slightly disrupts RAFP fingerprints.

Table 2: Fingerprint persistence under adaptive fine-tuning w/ rare-data injection (%).

Rare-data injection ratioGCG-QA Match ↑\uparrowRAFP Match ↑\uparrowRAFP-PPL Match ↑\uparrowUtility retention ↑\uparrow
0% (Standard SFT)6010080100
1%601008099.8
2%50908099.5
5%50908098.7
10%40807097.2

- Fingerprint matching rates use 10 fingerprints per method. Utility retention is normalized to the standard-SFT checkpoint. Rare-data injection ratio refers to the final training dataset size (50k).

W4: Missing related work

We thank the reviewer for identifying the missing citation to Gubri et al.'s TRAP. We will incorporate TRAP into the related work discussion and further discuss MEraser as an adaptive fingerprint erasure method. These works provide important context for distinguishing ordinary downstream fingerprint persistence from robustness against deliberate removal.

I thank the authors for their rebuttal. The results with the PPL aware optimization of fingerprints have addressed some of my concerns and I am raising my score. I would also appreciate if the authors could provide more context around the newly conducted fine-tuning experiments --- how were the rare-region samples constructed? Was any GCG done on them?

AuthorsOpenReview ↗

Clarification on Adaptive Rare-Region Fine-Tuning and Summary of Manuscript Revisions

Dear Reviewer A8e2,

We thank the reviewer for the follow-up question. In our adaptive fine-tuning experiment, we construct rare-region training pairs using the initialization procedure of RAFP (Section 4.1). Specifically, we first generate rare prompts using bottom-k sampling (k=2000) and then sample corresponding target responses conditioned on those prompts using the same bottom-k strategy. We do not perform GCG optimization on these injected pairs.

Intuitively, these samples are intended to increase the downstream training distribution's coverage of rare regions, rather than serve as optimized fingerprints themselves. They are generated independently of the secret verification fingerprints. This experiment therefore evaluates whether greater rare-region exposure alone can weaken fingerprint persistence.

Additional manuscript revisions

We have also incorporated the following updates into the revised paper, with all changes highlighted in blue:

  1. Adaptive fine-tuning evaluation (Section 6 Ablation, Table 4): We add the new experiments in the revised paper

  2. Perplexity-filtering evaluation (Section 6, Table 3a): We extend the comparison to include random triggers and GCG-QA with and without PPL regularization. The revised table reports fingerprint matching rates after applying the PPL filter, providing a clearer evaluation of the tradeoff between filter acceptance and fingerprint persistence.

  3. Additional related work (Section 2): We cite and mention TRAP (Gubri et al., 2024) and MEraser (Zhang et al., 2025) in our related work. We thank the reviewer for pointing out these two meaningful related works.

We sincerely thank the reviewer for the feedback and for reconsidering the score. Your suggestion helped us strengthen the paper.

Thanks for the clarification. In this case, for a more thorough adaptive attack I believe the attacker should match the fingerprint distribution, i.e. perform some steps of GCG on top of the rare region init. However I do not expect this to change the results much.

Reviewer_M5qJOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

RAFP introduces a "black-box fingerprinting" technique for tracing the provenance of large language models (LLMs). The method functions by identifying "prompt-response" pairs located within the "low-probability regions" of a base model; these prompts are then optimized across various system prompts and fine-tuned model variants to establish a fingerprint. This process requires no modifications to the model itself, with ownership verified via pre-committed hash values. The paper also provides a theoretical result that bounds the potential shift in a model's fingerprint following a single gradient update. The technique achieves perfect detection on base models and near-perfect results on the authors' own fine-tuned models—significantly outperforming training-based baselines. For publicly available derivative models, detection performance varies widely, while no false positives were observed on any of the unrelated models tested.

Critical strengths

Since RAFP searches for behavior that the model already exhibits, there is no need to modify the model itself. RAFP-generated fingerprints are also significantly more robust to fine-tuning compared to training-based baseline models, which is large enough to make their results convincing, even with the reporting inconsistencies discussed below. Ablation experiments are particularly helpful in demonstrating that optimizing against fine-tuned model variants is key to keeping fingerprints detectable at different stages of fine-tuning.

Critical weaknesses

Although perplexity filtering is explicitly included in the threat model, the experiment was conducted on only one model, and the main fingerprints in the paper were rejected by their own perplexity filter in all cases. More fundamentally, this contradiction seems to stem from the method itself: RAFP defines fingerprint rarity as a low probability under the base model, which is precisely the type of behavior that the perplexity filter is designed to detect. Considering the variant of the filter reduces rejection but does not eliminate it, resulting in lower detection accuracy, and it was evaluated only in a single ablation table.

Questions for authors
  1. What is the threshold used to determine whether two models belong to the same lineage? What false positive rate would this threshold produce in a larger pool of unrelated models?

  2. The key lemma seems to apply equally to common hints, since the proof does not assume fingerprints are rare. Therefore, theoretically or empirically, under what circumstances is rarity necessary?

2 comments
AuthorsOpenReview ↗

Addressing PPL-Filtering Vulnerability, Expanded 20 Negative Checkpoints for FPR Evaluation, and the Explanation of Rarity

We thank the reviewer for the constructive feedback and for recognizing RAFP's non-invasive fingerprint construction and robustness under downstream adaptation. We appreciate the concerns regarding perplexity filtering, verification reliability, and the theoretical role of rarity.

W1: Robustness against perplexity filtering

We agree that rare-region fingerprints can be vulnerable to perplexity-based filtering without explicit regularization. However, this vulnerability is not unique to RAFP only, since conventional GCG-based fingerprinting methods can also generate unnatural trigger sequences with high perplexity. We highlight our contribution is that rare-region optimization, combined with appropriate perplexity constraints could discover fingerprints that retain better model-lineage persistence under downstream model modification + practical filtering conditions. Rather than that our method are always inherently difficult to filter out.

We extend our experiments to include random triggers, GCG-QA, and a PPL-regularized GCG-QA baseline under the same perplexity filter. In our main paper Table 3(a), we only report matching rates of PPL-tuned fingerprint for base / SFT models without applying the filter during inference. In Table 1, we evaluate matching accuracy among fingerprints that successfully pass the filter during inference. Here, Base/SFT (Filtered) measures the matching rate among fingerprints that pass the PPL filter, and — indicates that no fingerprints survived filtering. The additional comparison demonstrates that conventional method GCG-QA suffer from similar filtering vulnerabilities and RAFP-PPL achieves a better tradeoff between filtering resistance and persistence.

Table 1: Fingerprint robustness under perplexity filtering (%).

MethodPPL Rejection ↓\downarrowBase (Filtered) ↑\uparrowSFT (Filtered) ↑\uparrow
Rand. triggers90 (9/10)0 (0/1)0 (0/1)
GCG-QA w/o PPL80 (8/10)100 (2/2)50 (1/2)
GCG-QA w/ PPL20 (2/10)87.5 (7/8)50 (4/8)
RAFP w/o PPL100 (10/10)——
RAFP w/ PPL (λ\lambda=0.1)20 (2/10)100 (8/8)87.5 (7/8)
RAFP w/ PPL (λ\lambda=0.3)10 (1/10)100 (9/9)66.7 (6/9)
RAFP w/ PPL (λ\lambda=0.5)10 (1/10)88.9 (8/9)66.7 (6/9)

- Values are percentages (counts). PPL rejection is rejected/total; Base/SFT matching is successful/surviving total.

Q1: Verification threshold and FPR

We clarify that our verification uses 10 fingerprint pairs, and a suspect model is classified as belonging to the protected lineage if at least one fingerprint reproduces its corresponding target response. Under this rule, main paper Table 2 identifies all nine related models and rejects all five unrelated models.

To further evaluate false-positive behavior, we expand negative-model pool to 20 more checkpoints outside the Llama 2 lineage, such as Mistra, , Qwen2.5, Gemma 2, DeepSeek, etc. The pool includes 14 pretrained base and six instruction-tuned checkpoints.

Table 2: Average fingerprint matching rates on unrelated checkpoints (%; lower is better).

Negative Model Group# CKPTsGCG-QA ↓\downarrowBase-only ↓\downarrowRAFP (Ours) ↓\downarrow
Pretrained base CKPTs1428.67.10
Chat/instruction-tuned CKPTs633.300
Overall203050

Results are average fingerprint matching rates across the unrelated checkpoints. The evaluation shows RAFP has the lowest FPR. Due to the character limit, we will report the complete per-checkpoint results in the revised paper.

Q2: The theoretical role and necessity of rarity.

We clarify that Lemma 1 explicitly requires (x^∈Rτ)(\hat{x}\in\mathcal R_\tau) (Lines 270–272) and invokes Assumption 1, which specifies that the rare-region probability mass under downstream fine-tuning satisfies (ρ(τ)≪1)(\rho(\tau)\ll1). Thus, rarity is an explicit condition of our theoretical analysis.

The theoretical motivation is that rarity reduces the likelihood that downstream fine-tuning samples come from the fingerprint's region, thereby limiting their direct contribution to fingerprint changes. Nevertheless, updates from common training examples may still affect the fingerprint through gradient alignment. We will make this additional condition explicit in the revised proof.

To empirically examine rarity, we compare fingerprints initialized by sampling from the top-2K (common), middle-ranked 2K (intermediate), and bottom-2K (rare) tokens based on the base model's token probabilities. All variants use identical optimization setting.

Table 3: Effect of prompt rarity on fingerprint persistence (%).

Prompt RegionBase ↑\uparrowSFT ↑\uparrowLoRA ↑\uparrow
Common1005060
Intermediate1007080
Rare (RAFP ours)100100100
AuthorsOpenReview ↗

Summary of Manuscript Revisions

Dear Reviewer M5qJ,

We have carefully addressed all three questions raised in our rebuttal and incorporated the new experimental results into the updated manuscript. All newly added content is highlighted in blue for your reference.

We would like to highlight the following updates in the paper revision:

  1. Robustness against perplexity filtering (Section 6, Table 3(a), p. 9). We extend our evaluation to include random triggers and GCG-QA with and without PPL regularization under the same filtering conditions. Unlike the original evaluation, we revised Table 3(a) to report fingerprint matching rates after applying the PPL filter. The results demonstrate that conventional GCG-based fingerprints also suffer from high rejection rates, while RAFP-PPL achieves a more favorable tradeoff between filter acceptance and fingerprint persistence.

  2. Expanded false-positive evaluation (Appendix A.5, Table 9, p. 20). To address concerns about our verification threshold and the limited number of negative models, we show the detailed results of evaluating 20 more unrelated checkpoints, including 14 pretrained base models and six instruction-tuned models. Under our one-of-ten fingerprint verification criterion, RAFP achieves 0% false-positive rate across all 20 additional checkpoints, providing stronger evidence of fingerprint uniqueness.

We hope these additional experiments and manuscript revisions address your concerns and further strengthen your assessment of our work. We would be happy to discuss any remaining questions or provide additional clarifications if needed. We sincerely appreciate your constructive feedback and consideration.

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

This paper studies the classic LLM fingerprinting problem. It proposes RAFP, a method that uses the base model and several adapted models to optimize rare prompt-response pairs (white-box access), while requiring only black-box access to the suspect model for verification. The paper provides a theoretical analysis of fingerprint persistence and evaluates the method on both self-finetuned SFT/LoRA models and publicly released downstream models, as well as under quantization. Ablations include prompt-template variations and filtering attacks.

Critical strengths

S1. The problem is well motivated and practically relevant, especially for model IP protection.

S2. The rare-region idea is intuitive and straightforward, and the author makes a good effort to justify it from both theoretical and empirical perspectives.

S3. The paper provides fairly comprehensive ablation, covering different prompt templates, decoding temperatures, and filtering attacks.

Critical weaknesses

W1. Insufficient experimental evaluation. This is my main concern. The paper provides some theoretical explanation for fingerprint stability under local gradient updates, but whether fingerprints survive substantial post-training still largely relies on empirical evidence. See below:

  1. Most experiments rely on relatively short SFT/LoRA training. There is a large 40% performance gap between these results and publicly released downstream models (Table 2).
  2. The method should be tested under more substantial continued training to see whether the fingerprints can persist, e.g., CodeLlama-7B (500B tokens) and Llemma-7B (700B tokens).
  3. Even on easy and well-studied cases like meditron-7b, RAFP performs poorly (?): only 2 out of 10 fingerprints match (20%), even lower than GCG-QA (60%). This needs more explanation.

W2. Theoretical assumptions need to be clarified. Lemma 1 requires the assumption that the expected gradient alignment between rare-region fingerprints and common training examples is close to zero. (Not stated) There is also a gap between the theoretical analysis of local gradient updates and the claimed persistence under prolonged training.

W3. White-box fingerprint construction needs to be clarified in both the introduction and limitations. Although verification is black-box, RAFP requires white-box access to the base model (and adapted models for optimization). This is not necessarily a major limitation, but it is different from the fully black-box setting, where both models are API-only, commonly discussed in existing literature. This distinction should be made clearer. Also, this reduces the significance of this work to some extent.

Questions for authors

Q1. What is the verification threshold in Figure 1, and how is it determined?

Q2. I would avoid using TPR/FPR here, as these terms can be misleading. The reported numbers are just fingerprint matching rates (or RAFP scores) over 10 prompts for each model, not the overall TPR/FPR of the proposed method across multiple model-pair tests.

Overall, I find the proposed method promising. One natural concern readers may have is whether the fingerprints can survive prolonged training. Previous white-box work like SeedPrints (ICLR 2026) shows that random, unnatural inputs (somewhat similar to rare-region prompts) can reveal intrinsic fingerprint signals that survive very long training. I think this may help address some concerns about fingerprint persistence. I strongly suspect the weaker results on models like meditron-7b and Orca-2-7b are due to insufficient prompt optimization, rather than the absence of persistent fingerprints. I would be happy to raise my score if the authors can show stronger results on these cases, as well as CodeLlama-7B and Llemma-7B.

4 comments
AuthorsOpenReview ↗

We thank the reviewer for the helpful suggestions. Additional evaluations and clarifications are provided for strengthening the evidence for RAFP's fingerprint persistence and provide a clearer understanding of its limitations under substantial downstream adaptation.

We thank the reviewer for the constructive feedback. We appreciate the suggestions regarding long-term fingerprint persistence and the limitations of optimization on challenging downstream models.

W1: Fingerprint persistence under prolonged training: We agree with the reviewer in pointing out SeedPrints (which we will cite), providing evidence that model-specific signals can persist through extensive training. While SeedPrints identifies initialization-dependent biases through internal model representations, RAFP constructs rare-region prompts as fingerprints that enable fingerprint verification on downstream models. Following the suggestion, we extend evaluation to two models, CodeLlama-7B and Llemma-7B, which undergo continued pretraining relative to Llama 2 family. Fingerprints are constructed without access to downstream checkpoints during optimization, allowing us to evaluate persistence under substantial model adaptation.

We also investigate the reviewer's hypothesis that weaker results on Meditron-7B and Orca-2-7B may arise from insufficient prompt optimization. As described in the paper, RAFP follows GCG optimization with a maximum of 500 optimization steps + multi-trial strategy that requires n successful triggers before accepting a fingerprint candidate. In our experiments, we use n = 20. We extend the optimization budget to 1K steps (2×) and 2K steps (4×), keeping the number of multi-trial strategies fixed.

In Table 1, we show the fingerprint matching rates (%) across downstream models and optimization settings

MethodStepsTrialsCodeLlama-7BLlemma-7BMeditron-7BOrca-2-7B
GCG-QA500140306030
RAFP (Original)5002050402040
RAFP-2× Steps1,0002060504040
RAFP-4× Steps2,0002070605050

Compared with GCG-QA, RAFP achieves better performance (70% and 60% fingerprint matching on CodeLlama-7B and Llemma-7B, respectively, with increased optimization), suggesting that most rare-region fingerprints remain transferable after prolonged pretraining. On Meditron-7B and Orca-2-7B, increasing the optimization budget improves matching from 20 --> 50% and 40 --> 50%, respectively.

W2: Theoretical assumptions. We thank the reviewer for raising this point. Our analysis is based on two complementary factors: the low probability mass of rare prompts under downstream fine-tuning (Assumption 1) and the limited gradient alignment between rare fingerprints and common training examples. The latter reflects the intuition that gradient contributions from common examples may partially cancel when projected onto rare-region fingerprint gradients. We agree that this condition should be stated more explicitly. We will formalize the expected alignment term in Lemma 1 and clarify its role in the likelihood-drift bound.

We would like to highlight that our theoretical analysis already extends beyond a single gradient update. Specifically, Corollary 2 (line 313-323) establishes a multi-step decoding-stability condition: for a fingerprint with initial decoding margin γ and a per-step drift bound B, the fingerprint remains top-ranked after t successive updates when γ>2tB\gamma \gt 2tB, with probability at least 1−2tδ1 − 2t \delta. We will clarify that this result requires the drift bound to hold throughout the update sequence. While the result does not guarantee stability under arbitrary prolonged training, it provides a theoretical characterization of how fingerprint persistence depends on the decoding margin and cumulative update drift. The new continued-pretraining experiments further examine persistence beyond the regimes covered by the theoretical conditions.

W3: White-box construction v.s. black-box verification. RAFP requires only white-box access to the base model. Access to adapted variants is optional. The verification requires only black-box query access to the suspected model. We will clarify this clearly.

Q1: Verification threshold. We use a set of 10 independently generated fingerprints for model verification. A suspect model is identified as belonging to the protected model's lineage if at least one of the 10 fingerprint prompts successfully triggers its corresponding target response. Thus, our verification threshold is one successful match out of 10 fingerprints (i.e., a 10% fingerprint matching rate). We will explicitly clarify this decision rule in Figure 1 and the experimental setup.

Q2: Evaluation terminology. We agree that per-fingerprint matching rates should be distinguished from model-level true-positive and false-positive rates. We will revise the terminology accordingly and define the verification score and model-level decision rule.

AuthorsOpenReview ↗

Addressed reviewer's questions and revised manuscripts

Dear Reviewer kcnA,

We have carefully addressed the concerns raised in the review (see above rebuttal), and incorporated theoretical and experimental revisions into the updated manuscript, with all changes highlighted in blue.

Specifically, we would like to highlight three major revisions in the paper:

  1. Strengthened theoretical analysis (Section 4.4, p. 6). We introduce Assumption 3 (Weak Cross-Region Gradient Alignment) and revise Lemma 1 and its proof to explicitly account for the expected gradient alignment between rare-region fingerprints and common downstream training examples. We also clarify the relationship between decoding margins and cumulative update drift in our stability analysis.

  2. Additional evaluation under prolonged training (Section 5, Table 2, p. 9). Following your suggestion, we extend our evaluation to CodeLlama-7B and Llemma-7B, two models derived from Llama 2 through continued pretraining. Our results show that RAFP persists under prolonged training.

  3. Expanded false-positive evaluation (Appendix A.5, Table 9, p. 20). We further added 20 unrelated model checkpoints, including 14 pretrained base models and six instruction-tuned models, to evaluate fingerprint uniqueness under the one-of-ten verification criterion. RAFP achieves 0% false-positive rate across all 20 additional checkpoints, further supporting its ability to distinguish related and unrelated model lineages.

We sincerely thank you again for your constructive feedback, and appreciate your time in reviewing our paper. We hope the revised analysis and additional empirical evidence help address your concerns, and we would be grateful if you could reconsider your assessment.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal and Manuscript Revisions

Dear Reviewer kcnA,

We sincerely thank you again for your constructive feedback. We would greatly appreciate it if you could review our updated manuscript and let us know whether the new results and clarifications have adequately addressed your concerns. We would be happy to discuss any remaining questions or provide further evidence if needed.

We hope these revisions strengthen your assessment of our work, and we would be grateful if you could reconsider your evaluation in light of the additional evidence.

Thank you again for your time and thoughtful consideration.

Thanks to the authors for the added results on prolonged training and the improved results on Meditron-7B and Orca-2-7B. My concerns have been resolved.

Re the threshold: since it’s picked by requiring at least 1 match among n fingerprinting prompts (n=10 here), it’s worth ablating n to show the sensitivity. This is an add-on, so I’m not waiting for another round of discussion and will increase my rating to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.