- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 2(+1)
Read review
Paper summary
The paper presents an attack on black-box LLM fingerprinting schemes. These schemes usually check the output of a target LLM on certain sets of queries, and use a classifier (learnt or heuristic) to determine if the outputs are from a specific LLM. The paper considers a threat model where an adversary can substitute a (cheaper or smaller) model in such a way that on the fingerprinting queries the outputs will still be classified as the target LLM. The paper achieves this by fine-tuning the smaller model to mimic the target LLM distribution on the distribution of fingerprint queries. The paper proves that this is feasible when the fingerprint auditor uses a small number of queries and a weak classifier. The paper also proposes a method to continually adapt the surrogate model to account for changing fingerprint query distributions, assuming that the adversary has knowledge about this. The attack is demonstrated to be successful to varying degrees on three existing schemes with a smaller model from the same target family as the surrogate.
Critical strengths
- I like the threat model adopted, which includes fingerprint spoofing, utility drop and economic gains. The latter two provide a practical incentive for an adversary to actually bypass fingerprinting schemes which is important.
- The theoretical results, while seemingly trivial, are nice to have written down. The takeaway in Lines 258-262 seems obvious in hindsight, however it does show that an adversary could bypass such fingerprinting schemes if they wanted. In fact I feel that this could also extend to stronger schemes like Gloaguen et al 2026.
- I liked the discussion around continual learning to efficiently adapt the adversarial model.
- The baseline attacks considered in the main results are nice. However, I would like to see one more baseline which is fine-tuning on general purpose data.
Critical weaknesses
Overall I feel that the paper is fine, but I do not see much significance of the work. This is mainly because of the assumption that the adversary knows the fingerprint query distribution.
- The paper claims that current fingerprinting methods do not consider an adversary finetuning the models to bypass fingerprinting mechanisms. This is categorcially false. Further, the key question asked is "can a malicious provider finetune their model to mimic a fingerprinted model only on the fingerprinted distribution"? This seems trivially possible especially if one knows the fingerprint query distribution. In my opinion this diminishes the significance of the work.
- Building on the first weakness, it seems extremely limiting that the adversary knows the fingerprint query distribution. The paper aims to justify this using Kerckoff's principle, however, even in this analogy usually the attacker knows broad strokes schemes and not the exact secrets (e.g. the secret key for encryption). I would argue that the exact fingerprint distributions are usually secret, and prior work (e.g. Nasery et al 2026) aim to disrupt the fingerprint verification without assuming access to the exact fingerprinting distribution (in their case this is a collection of discrete distributions). Indeed Nasery et al 2026 categorically do not rely on distinguishing fingerprint distributions in their sampling based attacks, applying the modified sampling uniformly to all queries.
- Missing prior work -- MEraser (Zhang et al 2025) which also aims to fine-tune models to make them forget fingerprints (though this mainly works against backdoor based fingerprints)
- In terms of the utility results, I would like to see more benchmarks which are generative in nature (like GSM8K), since fine-tuning based attacks often destroy generative capabilities much more than discriminative capabilities. From the current results, apart from Qwen based models GSM does suffer a non-trivial amount, which is concerning.
Refs Are Robust LLM Fingerprints Adversarially Robust?, Nasery et al, SaTML 2026 MEraser: An Effective Fingerprint Erasure Approach for Large Language Models, Zhang et al, ACL 2025
Questions for authors
- I know this might be out of scope for the current work, but how far can the assumption of "adversary knows the fingerprint distribution" be relaxed? I wonder if an extension is possible in a distributed setting where the attacker aims to learn the query distribution from samples?
- Why is economic cost a justified assumption? In my mind, given the amount of security investments at frontier labs, weight stealing is probably the domain of nation-state level adversaries (Mickens 2014). In that case, economic bottlenecks are not very concerning. Operationally, I wonder if removing the assumption of LoRA fine-tuning or few step fine-tuning changes the results of the paper by a lot.
- I also wonder how the attack success rate is measured, i.e. how many fingerprints was the model prompted with? For example, one could make the ASR arbitrarily low by increasing the number of queries. This could increase the FPR on the original models which needs to be calibrated.
- Could the authors comment on the necessary conditions for cross-model transfer to work?
- How were the auxiliary classifiers obtained? Are they even necessary? From table 7 in the appendix they usually do not have much impact on the results (e.g. on MET the ASR goes up by barely 3%).
9 comments
1st response to reviewer a7ke (2/3)
W3. MEraser.
Thank you for bringing MEraser [5] to our attention. We will cite it and discuss its relationship to our work. MEraser starts with a model carrying an implanted backdoor fingerprint and fine-tunes that model to remove the ownership signal. In our setting, the provider starts with a different, cheaper model, and success requires an auditor to accept it as a particular premium model. The two studies therefore address distinct risks of fine-tuning for model verification. We believe MEraser studies a completely different problem, and does not affect our novelty.
W4. Generative benchmark for model utility.
We added IFEval [6] to evaluate instruction following directly. We use strict prompt-level accuracy as the evaluation metric. The original weak model is the reference for measuring utility changes after adaptation.
| Model | Gemma-1.1 (%) | Qwen2 (%) | Phi-3 (%) |
|---|---|---|---|
| Original weak model | 30.13 | 26.25 | 48.98 |
| GhostPrint | 27.17 | 26.25 | 46.02 |
| Original strong model | 42.15 | 52.85 | 59.15 |
Overall, GhostPrint preserves much of the weak models’ utility across tasks. On IFEval, Qwen2 matches its original strict prompt accuracy, while the evaluated Gemma and Phi-3 checkpoints retain most of their respective baseline scores. In Table 2 of our paper, GSM8K is the only benchmark where a model loses a little bit more points. We believe more extensive hyperparameter tuning could further reduce such losses, and the empirical aspects of our work are sufficiently strong.
Q1. Can knowledge of the query distribution be relaxed to sample access?
Yes, to a point. The empirical procedure uses prompts sampled from the audit distribution. It does not evaluate a probability density or require the prompts that will appear in a future audit. Representative historical samples can therefore provide the required input even when the provider does not know the sampling rule explicitly.
The distributed setting raises a further question about representativeness. If several sites contribute samples, their local query distributions may differ from that of the customer conducting the audit. A natural extension is to evaluate a fixed model on held-out sites while varying the number of available samples per site, and to measure how audit acceptance changes with this shift. If the only observations are unlabeled mixtures of ordinary traffic and audit requests, recovering the audit distribution requires additional assumptions about that mixture. We see this as a worthwhile extension beyond the setting evaluated here.
Q2. Economic motivation and the role of LoRA.
We would like to clarify that this scenario requires no theft of frontier-model weights, and it is entirely irrelevant to "nation-state-level adversaries". Specifically, the adversary in Section 4 is a commercial inference provider. Consider a provider advertising an open-weight 7B model while serving a 1.5B model and charging the 7B price. Both models can be obtained legitimately. The customer’s question is whether the endpoint serves the model being billed, and the provider’s financial incentive comes from the recurring difference in serving costs.
Q3. Audit budget.
For LLMmap and LLM-idio, we follow the original paper's choice of queries and prompt source. As detailed in Appendix A.1, we evaluate all methods under the same audit budget on the same 100 held-out LLMmap probes and 2,000 held-out LLM-idio prompts, ensuring a fair comparison.
For MET, Appendix A.4 reports how ASR changes with the audit budget. Under the practical budget of one prompt and ten completions, GhostPrint achieves 60% ASR on Phi-3, while MET falsely rejects the genuine Phi-3 target in only 1 of 30 trials (Table 1).
Q4. Conditions for cross-model transfer.
Table 3 evaluates two cross-family directions, with a model adapted separately for each specified target. For Gemma-2B to Qwen2-7B and Qwen2-1.5B to Gemma-7B, LLM-idio ASR is 95.2% and 96.7%, LLMmap ASR is 0% and 6.0%, and MET ASR is 33.3% and 6.7%. These results establish that sharing a model family is unnecessary for success against some auditors. They also show why architecture alone is insufficient to predict the outcome.
Theorem 5.2 gives a more precise answer for its idealized linear verifier. With fixed weak-model features and an update rank at least as large as the verifier rank, exact matching of the auditor-observed target output is possible precisely when that output is linearly predictable from the weak-model features almost surely under the audit distribution. A practical classifier may accept a response even with nonzero matching error. We therefore present the theorem as a characterization of observable matching in that setting and Table 3 as evidence about actual audit decisions.
1st response to reviewer a7ke (3/3)
Q5. Construction and contribution of the auxiliary classifiers.
We train local surrogate classifiers on responses to training prompts for LLMmap and LLM-idio. For MET, we compute the test statistic directly (see Section 5.2). These scores are used only in the optional RAFT stage. Table 7 shows that RAFT improves LLM-idio ASR by 4.1–7.1 percentage points across the three model families. We agree that its gain on MET is modest (3.3 points).
Thank you again for your detailed questions and for suggesting the general-purpose fine-tuning and generative-utility evaluations. We would appreciate your reassessment of the paper in light of these new results and our clarifications.
References
[1] Dario Pasquini, Evgenios M. Kornaropoulos, Giuseppe Ateniese. "LLMmap: Fingerprinting for Large Language Models". USENIX Security 2025.
[2] Irena Gao, Percy Liang, Carlos Guestrin. "Model Equality Testing: Which Model Is This API Serving?". ICLR 2025.
[3] Mingjie Sun, Yida Yin, Zhiqiu Xu, J. Zico Kolter, Zhuang Liu. "Idiosyncrasies in Large Language Models". ICML 2025.
[4] Anshul Nasery, Edoardo Contente, Alkin Kaz, Pramod Viswanath, Sewoong Oh. "Are Robust LLM Fingerprints Adversarially Robust?". arXiv 2025.
[5] Jingxuan Zhang, Zhenhua Xu, Rui Hu, Wenpeng Xing, Xuhong Zhang, Meng Han. "MEraser: An Effective Fingerprint Erasure Approach for Large Language Models". ACL 2025.
[6] Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, Le Hou. "Instruction-Following Evaluation for Large Language Models". arXiv 2023.
1st response to reviewer a7ke (1/3)
Thank you for the detailed review. We sincerely appreciate your recognition of the practical motivation for our three objectives, the theoretical analysis, the continual-learning setting, and the baseline comparisons. Here are our responses to your concerns.
S4. Fine-tuning on general-purpose data
Thank you for suggesting this baseline. We fine-tuned Qwen2-1.5B on 2,000 Alpaca prompts under two conditions. The first uses the original Alpaca answers. The second uses answers generated by Qwen2-7B for the same 2,000 prompts. This comparison isolates the effect of changing the answer source while holding the prompts fixed.
Clearly, SFT with the original Alpaca answers does not improve the weak model’s attack success rate on any of the three auditors. Using Qwen2-7B answers does transfer some model-specific behavior, but the improvement is much smaller on LLMmap and MET. These results make general-purpose SFT an informative baseline. They show both what ordinary teacher imitation achieves and how much of the observed spoofing performance it leaves unexplained.
W1. Relationship to prior work and technical significance
(1) On characterization of prior work. LLMmap [1], MET [2], and LLM-idio [3] primarily study how to identify a model or detect changes in its outputs, while these works do not consider whether a malicious API provider exists. Specifically, LLMmap evaluates query-informed countermeasures, MET tests whether a fine-tuned model can be distinguished from its original checkpoint, and LLM-idiosyncrasies shows that students can inherit traits from their teachers. None of these evaluations asks whether a provider can deliberately fine-tune a cheaper model to pass an existing audit as a specified stronger model. To our knowledge, this targeted attack setting has also received limited attention in subsequent work on these LLM fingerprinting methods.
(2) On whether spoofing becomes trivial once the query distribution is known. Empirically, Table 3 shows that existing attacks and baselines have limited success in our known-distribution setting. For instance, Token Suppression from Nasery et al. [4] even achieves 0.0% ASR on the Gemma-1.1 + MET setting. Thus, even with the query distribution specified, success depends substantially on both the adaptation method and what the auditor measures.
The paper also theoretically establishes lower and upper bounds across linear models, causal attention, and Transformers. Specifically, Proposition D.3 gives a worst-case lower bound showing when a rank-constrained update cannot match the target across all inputs. Theorem 5.1 gives an upper bound on imitation error over the audit distribution in terms of the spectral tail of its input representations. Theorem 5.2 characterizes the minimum error on the outputs observable to a low-rank verifier. Our work is far beyond an empirical solution to this problem, and we also provide detailed theoretical analysis and deeper insights into this problem.
W2. Knowledge of the fingerprint query distribution
(1) The known-distribution setting is practical, and our assumption about query knowledge is stricter than the one studied in LLMmap. Consider a customer who uses the publicly released LLMmap tool to check which model an API serves. LLMmap publishes its default queries and explicitly studies a query-informed setting in which the service operator knows the exact queries used by the customer [1] (see Section 8.1). Our setting gives the provider less information, since it knows the distribution from which audit queries are drawn, but the actual queries used for evaluation are held out during training.
(2) We also study continual adaptation as new fingerprinting methods become available. In Section 6.4, we evaluate whether a provider can successively adapt the same model to additional fingerprints while retaining its ability to pass earlier audits. Thus, a provider can begin with the fingerprinting methods known at deployment and incorporate new ones as they emerge, allowing the same model to remain effective against an evolving set of audits.