acceptodds
Under review as a conference paper at ICLR 2027

Who Made This LLM? Black-Box Family Provenance for Large Language Models

Abstract

Frontier large language model (LLM) families such as Llama, Qwen and Mistral are valuable assets, and the ease of producing derivative models through fine-tuning, distillation or merging exposes them to theft, unauthorized reuse and false ownership claims. Identifying the family a deployed LLM descends from is therefore central to ownership auditing, but existing solutions fall short under realistic conditions. Watermarking requires proactive deployment and cannot be retrofitted to already-released models, white-box fingerprinting requires parameter access that API-served models do not expose, and extracting reliable family-level signals from black-box responses remains challenging under model modifications. We propose FamilyTrace, a black-box fingerprint for family-level LLM provenance that supports both closed-set attribution among known families and open-set rejection of unknown families. Our design starts from two observations: much of the knowledge a model learns in pre-training is encoded in its feed-forward layers, which hold most of its parameters, and typical post-training changes these weights only slightly. We therefore expect the models of a family to share similar word associations, and FamilyTrace captures them by asking a model to rank candidate words by their relation to an ambiguous word, first with little context and then with progressively more context. We evaluate FamilyTrace on two benchmarks covering six LLM families, a non-adversarial benchmark of 55 LLMs and an anti-provenance benchmark of 50 LLMs. FamilyTrace achieves 96.3% closed-set and 87.9% open-set accuracy on the non-adversarial benchmark, and 85.4% closed-set and 79.5% open-set accuracy on the anti-provenance benchmark, outperforming seven black-box fingerprinting baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Model Provenance, Black-Box Fingerprinting, Family-Level Attribution, Open-Set Recognition

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_DYybOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FamilyTrace, a black-box fingerprinting method for LLM family-level provenance identification. The key idea is that models from the same family share similar word association preferences even after model modifications. The method extracts fingerprints by asking models to rank candidate words under progressively informative contexts, and compares them with reference family fingerprints for closed-set and open-set identification. Experiments on 105 LLMs across six families demonstrate its effectiveness against various model modifications.

Critical strengths
  1. The paper is well-written, with a clear motivation and problem formulation.

  2. The fingerprinting design is simple and intuitive but effective. The experiments are comprehensive, covering various model modifications and comparisons with seven baselines.

Critical weaknesses
  1. The paper assumes that FFN layers store much of the pre-training knowledge and are only slightly modified during post-training, leading to similar word association preferences within the same model family. However, these assumptions are not directly validated.

  2. It is unclear whether FamilyTrace identifies model lineage or similarity in pre-training data. Independently trained LLMs sharing similar training corpora may exhibit similar word associations, potentially causing false attribution. Can increasing the number and diversity of anchors reduce this risk?

Questions for authors
  1. Could targeted fine-tuning on anchor words alter or remove the fingerprint?

  2. Could the authors further validate the assumptions about FFN layers and the effects of post-training?

  3. Could similar pre-training corpora lead to false attribution across different LLM families? Would more diverse anchors help?

  4. Missing reference on LLM fingerprinting: iSeal: Encrypted Fingerprinting for Reliable LLM Ownership Verification (Xiong et al., AAAI 2026).

3 comments
AuthorsOpenReview ↗

Response to Questions 1-2 and Weakness 1

We thank the reviewer for the careful assessment and the insightful questions.

Question 1: Could targeted fine-tuning on anchor words alter or remove the fingerprint?

Response: Targeted fine-tuning can alter or remove the fingerprint if the attacker knows the anchor words or the query list. Section 5 simulates such an attack by tampering with the rankings of some anchors. Since the similarity to a family is the median over the six anchors, CS Acc. on the non-adversarial benchmark stays at 0.826 or above when two anchors are tampered with, but attribution fails with three or more (Table 11). This simulation corresponds to the case where fine-tuning fully controls the rankings of the targeted anchors. Such an attack is outside our threat model, which assumes that the provider does not adapt its responses to fingerprinting queries (Section 3.2), and the same risk applies to fingerprints that rely on fixed queries. The auditor can reduce this risk by keeping query lists private and generating a new list for each audit, so that the attacker does not know which anchors to target.

Weakness 1: The paper assumes that FFN layers store much of the pre-training knowledge and are only slightly modified during post-training, leading to similar word association preferences within the same model family. However, these assumptions are not directly validated.

Question 2: Could the authors further validate the assumptions about FFN layers and the effects of post-training?

Response: The two observations about FFN layers and post-training are drawn from prior work (Section 1). They motivate our hypothesis that the models of a family largely share their word association preferences, and Section 5 tests this hypothesis through its predictions. In a controlled experiment with three base models and three fine-tuning datasets (Table 15), models fine-tuned from the same base model on different data are more similar to each other than models fine-tuned on the same data from different base models (mean pairwise similarity of 0.241 versus -0.005, with a p-value below 10−410^{-4} in a permutation test). Queries with common-sense orderings also carry less family information.

During the rebuttal, we also measured on the models of our benchmarks how large FFN layers are and how much post-training changes the weights. FFN layers hold 63.4% to 77.8% of the parameters of the six base models. For 74 fine-tuned models of our benchmarks, all 50 anti-provenance models and 24 non-adversarial models, we compared the weights with those of the parent model, using the Frobenius norm of the update relative to that of the parent weights, computed for each weight matrix and averaged in proportion to the matrix sizes. Table R5 reports the change for each family.

Table R5. Relative weight change of 74 fine-tuned models, compared with their parent models.

FamilyModelsMedian changeMaximum changeMaximum FFN change
Gemma-2-2B160.29%0.63%0.86%
Llama-3.1-8B170.13%0.37%0.44%
Mistral-7B100.65%1.40%1.52%
Qwen-2.5-7B120.07%1.85%2.12%
Qwen-2.5-14B140.10%0.18%0.21%
TinyLlama-1.1B50.28%0.64%0.72%
All740.14%1.85%2.12%

As shown in Table R5, post-training changes the weights of every family only slightly, by at most 2.12% for the FFN weights. Of the 74 models, 23 do not change FFN weights at all, since they are LoRA adapters that only update attention weights.

The size of the change also relates to the fingerprint. On the anti-provenance benchmark, a larger change goes with a lower own-family similarity (Spearman correlation of -0.827 over 50 models). On the non-adversarial benchmark, all 24 models keep an own-family similarity between 0.913 and 0.997. These results support the observation about post-training and indicate that larger changes of the weights preserve less of the fingerprint.

AuthorsOpenReview ↗

Response to Questions 3-4 and Weakness 2

Weakness 2: It is unclear whether FamilyTrace identifies model lineage or similarity in pre-training data. Independently trained LLMs sharing similar training corpora may exhibit similar word associations, potentially causing false attribution. Can increasing the number and diversity of anchors reduce this risk?

Question 3: Could similar pre-training corpora lead to false attribution across different LLM families? Would more diverse anchors help?

Response: In our setting, a family is one pre-trained base checkpoint and the models derived from it (Section 3). If the fingerprint reflected the pre-training corpus instead of the checkpoint, models of the same series would be hard to separate. Qwen-2.5-7B and Qwen-2.5-14B belong to a series pre-trained on one corpus of 18 trillion tokens (Qwen2.5 technical report, arXiv:2412.15115). They are separate families in the main benchmarks, and models of one are attributed to the other in at most 2.9% of the cases. This is a harder case than independently trained models with similar corpora, since the two models also come from one developer and share a unified vocabulary. Across the five Qwen-2.5 sizes from 0.5B to 32B in Appendix B.11, CS Acc. is 0.940. The fingerprint therefore follows the checkpoint rather than the corpus alone.

The risk the reviewer describes appears in the open-set setting. When Qwen-2.5-32B is held out as unknown, 20% of its models are accepted as another Qwen-2.5 size (TPR 0.800, Appendix B.11). These errors stay within one developer and series. We will state this boundary in the limitations, and the results below show that more anchors reduce it.

More anchors also improve attribution. In Appendix B.8, increasing the number of anchors from 3 to 7 raises CS Acc. from 0.882 to 0.939 on the non-adversarial benchmark and from 0.836 to 0.901 on the anti-provenance benchmark. To examine the risk directly, during the rebuttal we extended the default query list on the five Qwen-2.5 sizes of Appendix B.11 with query lists of Appendix B.12 that change the anchor words, which adds anchors from new semantic groups.

Table R6. Results on the five Qwen-2.5 sizes when the default query list is extended with lists of new anchors, averaged over 20 random draws of the added lists.

Query listsQueriesCS Acc.OS Acc.TPRFPRAUROC
2480.9540.8750.9760.1510.968
4960.9870.8860.9930.1410.994
81921.0000.8831.0000.1461.000

As shown in Table R6, TPR, the proportion of unknown models correctly rejected, and AUROC increase with the number of anchors and both reach 1.000 with 192 queries. More and more diverse anchors therefore reduce the risk that a family of the same series is accepted as a known one.

Question 4: Missing reference on LLM fingerprinting: iSeal: Encrypted Fingerprinting for Reliable LLM Ownership Verification (Xiong et al., AAAI 2026).

Response: Thank you for the pointer. iSeal (Xiong et al., AAAI 2026) embeds an encrypted fingerprint into a model before release and uses it to verify the ownership of that specific model. FamilyTrace addresses a complementary setting. It requires no change to the model and attributes already released models at the family level, including models into which no fingerprint was embedded. We will add iSeal to the related work in the revision.

AuthorsOpenReview ↗

Follow-up to Reviewer DYyb

Dear Reviewer DYyb,

Thank you again for your review. We have posted our responses to your comments, which discuss whether targeted fine-tuning on anchor words can alter the fingerprint (Q1), validate the assumptions about FFN layers and the effects of post-training (W1 and Q2), examine model lineage versus similarity in pre-training data and the effect of more diverse anchors (W2 and Q3), and add the missing reference to iSeal (Q4).

We would be very happy to discuss any remaining questions. We sincerely appreciate your time and effort, and we hope that our responses are helpful for your evaluation of our work.

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

This paper presents FamilyTrace, a black-box method for identifying the base-model family of a modified LLM. It compares candidate-word rankings under progressively informative contexts with reference family profiles to attribute known families or reject unknown ones, reporting stronger overall results than seven evaluated baselines.

Critical strengths
  • The paper studies an important and interesting problem.
  • The method requires only 24 text queries per model and supports both known-family attribution and unknown-family rejection.
Critical weaknesses
  • The contribution beyond closely related work is unclear.
  • The most relevant baseline is missing without explanation.
  • The effect of anchor-word exposure during fine-tuning is untested.
Questions for authors
  • My main concern is that the Introduction does not clearly establish FamilyTrace’s contribution relative to closely related black-box provenance work. It argues that existing methods rely on “surface-level signals such as stylistic patterns or trigger responses” that are vulnerable to deployment-time changes. This characterization does not adequately account for ErrorTrace, which identifies model families using shared error patterns. Although staged word-association rankings are a distinct design, the paper does not explain what limitation of such family-level signals they address. As a result, the motivation overgeneralizes prior work and leaves the incremental contribution unclear.

  • The selection of experimental baselines is not sufficiently justified. The paper compares FamilyTrace with seven black-box methods under a 24-query budget, but does not explain why this set covers the most relevant prior work. In particular, it identifies ErrorTrace as a closely related method for black-box model-family attribution, yet omits it from the closed-set comparison without explaining why a comparison is infeasible. This makes it difficult to assess FamilyTrace’s incremental advantage over the closest prior work. The authors should explain why the methods cannot be meaningfully compared.

  • The robustness of FamilyTrace when its anchor words occur frequently in fine-tuning data remains unclear. Such data could shift the semantic associations on which the fingerprint relies. Although Appendix B.14 tests different fine-tuning datasets, it does not examine anchor-word exposure or whether changes in their usage affect attribution accuracy.

6 comments
AuthorsOpenReview ↗

Response to Question 1

We thank the reviewer for the precise and constructive assessment.

Question 1: My main concern is that the Introduction does not clearly establish FamilyTrace's contribution relative to closely related black-box provenance work. It argues that existing methods rely on "surface-level signals such as stylistic patterns or trigger responses" that are vulnerable to deployment-time changes. This characterization does not adequately account for ErrorTrace, which identifies model families using shared error patterns. Although staged word-association rankings are a distinct design, the paper does not explain what limitation of such family-level signals they address. As a result, the motivation overgeneralizes prior work and leaves the incremental contribution unclear.

Response: We agree that the sentence in the Introduction overgeneralizes prior work and does not account for ErrorTrace, which uses error patterns shared within a family. Both methods use family-level signals, and the difference lies in how the signal is obtained. ErrorTrace finds it among queries from classification benchmarks that models happen to answer incorrectly, while FamilyTrace elicits it by design with queries that have several reasonable answers. This design addresses two limitations of error patterns.

  • Query cost. In ErrorTrace, a query contributes to the fingerprint only when a model answers it incorrectly, and mainly when the error is specific to one family, so the informative queries have to be collected from a large pool. Auditing one suspect model uses 10K queries (Section 7.2 in their paper). In FamilyTrace, every query returns a full ranking of the candidates that reflects the word association preferences of the model, and 24 queries per model are sufficient. New query lists can be generated by the same design rule with stable accuracy (100 query lists, our Appendix B.12).
  • Family granularity. ErrorTrace computes its error patterns over the models of a family, and its families group models of different versions and sizes, for example a Qwen family of eleven models. Its authors report misidentification when the models of a family differ in scale (Section 7.1 in their paper). FamilyTrace extracts the fingerprint of each model separately, so each family can be identified by one pre-trained base checkpoint, and the modified variants of five Qwen-2.5 families from 0.5B to 32B are attributed with 0.940 CS Acc. (our Appendix B.11).

The contribution of FamilyTrace relative to ErrorTrace is therefore to show that a family-level signal can be elicited by design, with a small number of queries and at the level of single base checkpoints. We will revise the statement in the Introduction accordingly.

AuthorsOpenReview ↗

Response to Questions 2-3

Question 2: The selection of experimental baselines is not sufficiently justified. The paper compares FamilyTrace with seven black-box methods under a 24-query budget, but does not explain why this set covers the most relevant prior work. In particular, it identifies ErrorTrace as a closely related method for black-box model-family attribution, yet omits it from the closed-set comparison without explaining why a comparison is infeasible. This makes it difficult to assess FamilyTrace's incremental advantage over the closest prior work. The authors should explain why the methods cannot be meaningfully compared.

Response: ErrorTrace is closely related to our setting. Our experimental comparison, however, is organized around the query budget. A small query budget is one of our design goals (Section 3.2), so all methods are evaluated under the same budget of 24 queries.

Under this budget, the seven baselines cover the main approaches in prior black-box fingerprinting: optimized prefix-based queries (ProFLingo), trigger-pattern and knowledge-level fingerprints (DuFFin), response embeddings (SEF), query-response traces (LLMmap), output distributions (MET), prompt perturbations (ZEROPRINT) and LLM-based judging (HideNSeek).

ErrorTrace is designed for a much larger budget. Its 10K queries per suspect model are about 400 times our budget, and its authors report that the accuracy decreases to 0.59 when 70% of the test queries are removed to reduce the audit cost (Section 7.2 in their paper). ErrorTrace therefore cannot be run as designed at 24 queries, and we did not include it in the comparison.

Question 3: The robustness of FamilyTrace when its anchor words occur frequently in fine-tuning data remains unclear. Such data could shift the semantic associations on which the fingerprint relies. Although Appendix B.14 tests different fine-tuning datasets, it does not examine anchor-word exposure or whether changes in their usage affect attribution accuracy.

Response: In the three fine-tuning datasets of Appendix B.14, each of the six anchors of the default query list occurs on average at most 47 times per 3,000 examples in Alpaca and MetaMathQA and 54 to 399 times in UltraChat. To examine anchor-word exposure directly, during the rebuttal we selected 3,000 examples from Alpaca and UltraChat in which each of the six anchors occurs 519 to 2,078 times, together with 3,000 control examples from the same sources that contain none of them. We fine-tuned the instruct model of each family on each set with the LoRA setting of Appendix B.14 and two seeds, and evaluated the 12 models of each set on the default query list.

Table R4. Own-family similarity for each anchor of the default query list, after fine-tuning with and without the anchors.

Fine-tuning dataaccountbankcapitalchannelshorestream
None (instruct models)0.9950.9840.9880.9890.9870.945
Control data without the anchors0.8960.8050.9720.9260.8720.875
Data with frequent anchors0.9770.7860.9710.9520.9670.930

As shown in Table R4, for every anchor, the similarity of the models fine-tuned on the data with frequent anchors is at the same level as that of the control models. For LoRA fine-tuning at the scale of Appendix B.14, frequent exposure therefore does not shift the rankings beyond the effect of fine-tuning in general. All 12 of these models are correctly attributed to their own family.

Thank you for the response and the additional experiment. Your response addresses most of my concerns. I have one remaining question about the frequent-anchor set. Could you report how many training examples contain multiple anchor words, including within the same sentence? Could you also compare the models’ anchor-word rankings before and after fine-tuning in both groups? This would help clarify which semantic associations are preserved and whether frequent exposure changes which candidate words the models prefer.

Reviewer_jCktOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Family Trace, a black-box family-level provenance method for LLMs. Based on the hypothesis that pre-trained associative knowledge stored in FFN layers remains largely preserved after post-training, Family Trace uses a 24-query protocol (6 polysemous anchors ×\times 4 progressive context stages) requiring the model to rank candidate words by relatedness. It constructs stage-wise preference profiles and compares target models against known family profiles using the median Pearson correlation over anchors. Evaluated on two benchmarks covering 6 families (55 non-adversarial models and 50 anti-provenance models), it outperforms 7 black-box baselines.

Critical strengths
  • Clever and efficient query design: Requiring only 24 text-level forced-choice ranking queries without needing internal logits or white-box parameter access makes it highly practical for third-party API auditing.
  • Rigorous theoretical and empirical analysis: Proposition 1 provides a clean interpretation of family profile similarity as normalized rank agreement. The paper also includes thorough auxiliary evaluations, such as 100 query-list variants, random vs. common-sense query comparisons, and a variance decomposition separating base models from fine-tuning data.
Critical weaknesses
  • Reference library contamination / Near-duplicate bias in Leave-One-Out evaluation: In the evaluation protocol, leaving out only the target model LtL_t means that all other close variants (e.g., 11 sampling variants, 12 system prompt variants, and other LoRA checkpoints sharing the exact same base model) remain in the reference library Fk′\mathcal{F}_k'. Appendix B.14 (Table 15) reveals that when models fine-tuned on the same dataset or other seeds are removed from the library, closed-set attribution drops drastically (e.g., dropping from 96.3% down to 55.6% when removing models from the target's dataset). In real-world audits, an auditor typically has only the official base/instruct model, not a dense cluster of specialized fine-tuned variants.
  • Lack of clarity on format compliance and parsing failures: The prompt design strictly mandates a specific output format (ANSWER: word > word > word...) without ties. The paper does not discuss how parsing failures, missing words, or malformed outputs from smaller or adversarially modified models (e.g., N-SFT models) were handled during evaluation.
  • Counter-intuitive anomaly in anti-provenance results: In Table 4 (Appendix B.5), injecting Laplace embedding noise during single/multi-teacher imitation (N-STI and N-MTI) unexpectedly improves closed-set accuracy compared to clean imitation (jumping from ~0.83 to 0.950). This counter-intuitive jump lacks a convincing explanation.
  • Vulnerability to input-level filtering or prompt defenses: Because the queries follow a rigid structural and textual template, API providers or defense wrappers could easily detect and disrupt them (e.g., by randomizing candidate order or detecting sorting tasks).
Questions for authors

see weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Weakness 1

We thank the reviewer for the careful reading and the constructive comments.

Weakness 1: Reference library contamination / Near-duplicate bias in Leave-One-Out evaluation: In the evaluation protocol, leaving out only the target model LtL_t means that all other close variants (e.g., 11 sampling variants, 12 system prompt variants, and other LoRA checkpoints sharing the exact same base model) remain in the reference library FkF_k. Appendix B.14 (Table 15) reveals that when models fine-tuned on the same dataset or other seeds are removed from the library, closed-set attribution drops drastically (e.g., dropping from 96.3% down to 55.6% when removing models from the target's dataset). In real-world audits, an auditor typically has only the official base/instruct model, not a dense cluster of specialized fine-tuned variants.

Response: An auditor who holds only the official base or instruct model of each family is a realistic scenario. During the rebuttal, we evaluated this scenario on both benchmarks with a reference library that contains only the instruct model of each family, averaged over the 100 query lists.

Table R1. CS Acc. with different reference libraries, as mean ± standard deviation over the 100 query lists.

Reference libraryNon-adversarial CS Acc.Anti-provenance CS Acc.
Leave-one-out (paper)0.963 ± 0.0220.854 ± 0.035
One instruct model per family0.942 ± 0.0310.825 ± 0.052

(On the non-adversarial benchmark, the six instruct models serve as references in the second row and are not evaluated as targets, which leaves 49 target models. For these 49 models, CS Acc. under leave-one-out is 0.960.)

As shown in Table R1, with only the instruct model of each family as the reference, the library contains no sampling, system-prompt or LoRA variants of the target, and CS Acc. is 0.942 and 0.825, close to 0.963 and 0.854 under the leave-one-out protocol.

Table 15 (Appendix B.14) is a robustness check under a stress-test scenario. Three base models are fine-tuned on the same three datasets with identical settings, so models from different families share their training data, and the data can pull a target toward another family. When the other seeds of the target's configuration (same base model and dataset) are removed, only the competing families keep models fine-tuned on the target's dataset, and CS Acc. is 0.556. When all models fine-tuned on the target's dataset are removed from every family, CS Acc. is 0.667. Both remain above the 0.333 expected by chance for three families.

Together, these results indicate that the attribution relies on a signal shared across the family instead of on near-duplicate variants of the target. We will include Table R1 as an additional evaluation in the revision.

AuthorsOpenReview ↗

Response to Weakness 2

Weakness 2: Lack of clarity on format compliance and parsing failures: The prompt design strictly mandates a specific output format (ANSWER: word > word > word...) without ties. The paper does not discuss how parsing failures, missing words, or malformed outputs from smaller or adversarially modified models (e.g., N-SFT models) were handled during evaluation.

Response: We describe below how outputs that do not follow the required format are handled and how often they occur. For each query, the ranking is obtained as follows.

  1. If the response does not contain a valid ranking of all five candidates in the required format, the query is sampled again with a different temperature, up to 10 attempts.
  2. If no attempt gives a valid ranking, the candidates are ranked by the order in which they first appear in the last response, which reflects the order in which the model generates them.
  3. Candidates that do not appear in that response are placed after those that appear, in the order in which they are listed in the query.
  4. If no candidate appears, the query is counted as a parsing failure, and the candidates keep the order in which they are listed in the query.

Table R2 reports, over the 100 query lists, the share of queries whose ranking is a valid ranking within the 10 attempts (step 1), is taken from the order of appearance (steps 2 and 3), or is a parsing failure (step 4).

Table R2. Fraction of queries with each parsing outcome, over the 100 query lists.

ModelsValid ranking (step 1)Order of appearance (steps 2 and 3)Parsing failure (step 4)
Non-adversarial benchmark
Gemma-2-2B0.8910.1090.000
Llama-3.1-8B0.8930.1060.000
Mistral-7B0.7970.1740.029
Qwen-2.5-7B0.8760.1140.010
Qwen-2.5-14B0.8410.1590.000
TinyLlama-1.1B0.4460.5070.047
Anti-provenance benchmark
N-SFT0.7770.2120.012

As shown in Table R2, all families except TinyLlama-1.1B give a valid ranking in 0.797 to 0.893 of the queries, and parsing failures stay below 0.050. TinyLlama-1.1B follows the format least often, but CS Acc. for its models is still 0.978. On the anti-provenance benchmark, the N-SFT models give a valid ranking in 0.777 of the queries with a parsing failure rate of 0.012, close to the non-adversarial models.

AuthorsOpenReview ↗

Response to Weaknesses 3-4

Weakness 3: Counter-intuitive anomaly in anti-provenance results: In Table 4 (Appendix B.5), injecting Laplace embedding noise during single/multi-teacher imitation (N-STI and N-MTI) unexpectedly improves closed-set accuracy compared to clean imitation (jumping from ~0.83 to 0.950). This counter-intuitive jump lacks a convincing explanation.

Response: During the rebuttal, we re-examined the attribution of every model in the anti-provenance benchmark. The gap between imitation with and without noise comes from the two Mistral-7B students in each imitation setting, while the students of the other families are almost unaffected by the noise.

Table R3. CS Acc. for each attack on the anti-provenance benchmark, with and without the Mistral-7B models.

AttackCS Acc. (all models)CS Acc. (without Mistral-7B models)Own-family similarity of the Mistral-7B models
STI0.8230.9680.403
N-STI0.9500.9650.800
MTI0.8390.9810.440
N-MTI0.9500.9690.803
N-SFT0.7100.6700.787

As shown in Table R3, without the Mistral-7B students, the noise makes almost no difference to CS Acc. We also compared the similarity of these Mistral-7B students to their own family. STI and MTI lower it to 0.403 and 0.440, while with noise it stays at about 0.80, the same level as Noisy SFT (N-SFT, 0.787). A possible explanation is that the embedding noise weakens the imitation itself, so the students adopt fewer of the teacher's preferences.

Weakness 4: Vulnerability to input-level filtering or prompt defenses: Because the queries follow a rigid structural and textual template, API providers or defense wrappers could easily detect and disrupt them (e.g., by randomizing candidate order or detecting sorting tasks).

Response: Query detection is a relevant concern for any fingerprint that relies on active queries. Our threat model (Section 3.2) assumes that the provider does not detect fingerprinting queries or adapt its responses to them, and we discuss below how FamilyTrace behaves when this assumption is relaxed.

Disrupting the queries requires detecting them first. The queries do not follow a single template. The 100 query lists in Appendix B.12 vary the wording of the sentences, the anchor and candidate words, and the ranking instruction and output format, and they reach similar CS and OS Acc. The auditor can therefore use a new list in each audit, and a provider would need to flag ranking requests in general, so randomizing their answers would also affect ordinary users who ask such questions. If a provider alters the rankings for some of the anchors, the median over anchors limits the effect, and CS Acc. on the non-adversarial benchmark is still 0.903 when two of the six anchors are tampered with by reversal.

AuthorsOpenReview ↗

Follow-up to Reviewer jCkt

Dear Reviewer jCkt,

Thank you again for your review. We have posted our responses to your comments, which evaluate a reference library with only one instruct model per family (W1), clarify format compliance and parsing (W2), trace the accuracy change under noisy imitation models (W3), and discuss how FamilyTrace behaves when queries may be detected (W4).

We would be very happy to discuss any remaining questions. We sincerely appreciate your time and effort, and we hope that our responses are helpful for your evaluation of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.