acceptodds
Withdrawn from ICLR 2027

SinkMark: Robust Model Fingerprinting via Attention-Sink-Induced Hallucinations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large Vision-Language Models (LVLMs) are increasingly released and adapted for downstream applications, making it challenging to verify whether a suspicious black-box model is derived from a copyrighted checkpoint. In this work, we propose SinkMark, a robust model fingerprinting framework that turns model hallucinations from an undesirable failure mode into a behavioral ownership signal. Our key idea is to construct private image-text probes that induce a predefined target hallucination in the source model and its derivatives, while remaining inactive on unrelated models. To achieve this, SinkMark dynamically identifies a shared attention sink across multiple decoder layers, promotes cross-layer agreement and attention concentration at the sink, and binds its hidden representation to the target hallucination semantics. We jointly optimize a bounded image perturbation and a natural-language suffix using PGD and GCG, respectively. To improve persistence under downstream model modifications, we further introduce robust min–max optimization with temporary adversarial parameter updates and predecessor-guided target induction. Importantly, SinkMark does not permanently modify the released model and requires only black-box access during ownership verification. Experiments demonstrate that SinkMark provides reliable and model-specific target hallucinations, while remaining robust to fine-tuning, pruning, and quantization.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Vision-Language Models, Model Fingerprinting, Black-box Ownership Verification, Attention Sink, Robust Hallucination

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qKhNOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
1
Presentation
2(−1)
Soundness
2
Read review
Paper summary

SinkMark uses attention-sink-induced hallucinations as behavioral fingerprints to identify vision-language models and their modified derivatives through black-box queries.

Critical strengths
  1. Novel fingerprinting mechanism: Turns attention-sink-induced hallucinations into model-specific signals for black-box ownership verification.

  2. Robustness without permanent model changes: Preserves fingerprint signals across several fine-tuning and compression settings while leaving the released model unchanged.

Critical weaknesses
  1. White-box probe construction: Probe construction requires access to model internals and gradients. Although verification is black-box, probe construction with only API access to the protected model is not demonstrated.

  2. Limited evaluation: Only two source models are tested, with insufficient analysis of verification query budgets and false-positive rates.

Questions for authors

How to apply the method on models without clear or stable attention sinks?

2 comments
AuthorsOpenReview ↗

Response to Reviewer: Access Assumptions, Evaluation, and Applicability

Thank you for the constructive comments.

Q-1. White-box probe construction: Probe construction requires access to model internals and gradients. Although verification is black-box, probe construction with only API access to the protected model is not demonstrated.

RESPONSE: We agree: SinkMark requires white-box source access for construction and does not demonstrate API-only construction. The intended user is a publisher with access to the protected checkpoint, checking whether a suspicious black-box service derives from it.

Construction uses source hidden states, attention weights, and gradients to optimize private image-text probes (Sections 4.1-4.2). Temporary parameter updates are discarded, leaving the released model unchanged. Verification uses only the suspicious model's input-output interface (Section 3.2).

Section 6 explicitly states this limitation. Commercial models in Appendix A.10 are unrelated negative controls, not evidence of API-only construction. Extending construction to API-only access remains future work.

Q-2. Limited evaluation: Only two source models are tested, with insufficient analysis of verification query budgets and false-positive rates.

RESPONSE: We add a third source model, query-budget experiments, and empirical false-positive rates on four unrelated models.

Additional model. Appendix A.9/Table 7 evaluates MiniGPT-4 with LoRA and full fine-tuning on four tasks. SinkMark has the highest TMR in five of eight settings; AGDI or SIF leads in the remaining three. All results are reproduced below.

Query budgets. Appendix A.11/Figure 17 uses a fixed pool of 200 probes, a fixed seed, and 100 random-subset trials per budget, Q∈{1,5,10,20}Q \in \lbrace 1,5,10,20 \rbrace. A trial is positive when its TMR meets ηQ=1/Q\eta_Q=1/Q. The source has 100% detection at all budgets. Across eight derivatives, detection ranges from 17-68% at Q=1 and 73-100% at Q=5; seven reach 100% at Q=10, and all eight at Q=20.

False positives. The same protocol evaluates Qwen2.5-VL-3B-Instruct, Qwen2.5-VL-7B-Instruct, MiniGPT-4-Vicuna7B, and InstructBLIP-Vicuna7B. Each yields 0/100 positive trials at every budget, complementing the unrelated-model TMR results reported in Table 1. These are empirical rates for these models and this fixed pool, not guarantees for arbitrary unrelated models.

MiniGPT-4 TMR (manuscript Table 7). L/F denote LoRA/full fine-tuning. Task order: MathV, PaintingForm, TextVQA, V7W. Bold marks column maxima.

MethodL-MathVL-PaintingFormL-TextVQAL-V7WF-MathVF-PaintingFormF-TextVQAF-V7W
Ordinary1%4%9%1%1%9%4%7%
IF9%17%12%11%2%19%11%9%
RNA2%15%9%6%8%9%12%9%
PLA17%27%17%18%34%46%23%27%
CroPA33%13%21%29%0%10%1%0%
DIM21%29%19%27%8%19%15%6%
ProFLingo22%34%19%11%22%19%25%10%
AGDI48%54%29%50%34%66%31%61%
SIF44%50%19%40%37%59%32%40%
SinkMark49%56%28%52%42%67%29%54%

Figure location: Revised manuscript, Appendix A.11, Figure 17, PDF p. 22. Left: source/derivative detection rates. Right: four unrelated-model FPR curves overlapping at 0%. Each budget uses 100 trials; the threshold is ηQ=1/Q\eta_Q=1/Q.

Q-3. How to apply the method on models without clear or stable attention sinks?

RESPONSE: SinkMark selects a candidate and attempts to induce sink-like behavior; it does not require an initially strong sink. Section 4.1 uses cross-layer hidden-state similarity (Eq. 8), following Mirage, to locate a potential aggregation position. This is a selection heuristic, not an attention-concentration measurement. The attention objective encourages attention to the position, while consistency aligns cross-layer candidate distributions (Eqs. 13-15). Binding and post-sink induction associate it with the target hallucination (Eqs. 9-10).

Successful induction is not guaranteed: construction may fail if optimization cannot establish attention concentration and target association. Section 6 states that effectiveness on architectures where these patterns cannot be induced, or with substantially different attention mechanisms, remains unverified. This is an applicability limitation, not architecture-independent robustness.

Reviewer_qKhNOpenReview ↗

Follow up comment

Several critical concerns remain unresolved and need to be addressed before I can reconsider my recommendation.

  1. The additional model does not adequately address the concern about evaluation coverage. Your new model MiniGPT-4 dates back to 2023. While the manuscript also evaluates Qwen2.5-VL, adding another legacy model does little to establish relevance to current VLMs. For an ICLR 2027 submission, can the authors evaluate at least one representative model released in 2026, including its modified derivatives? Attention-sink behavior, visual processing, and susceptibility to adversarial perturbations can differ substantially across model generations. These differences are particularly relevant here because the proposed method explicitly depends on inducing specific internal attention behavior. The response acknowledges that induction may fail, but provides no evidence establishing when or why. Increasing the model count does not, by itself, resolve the applicability concern.

  2. The new query-budget experiment does not sufficiently establish reliable ownership verification. The reported zero false positives come from repeatedly subsampling a fixed pool of 200 probes on only four negative-control models. These trials characterize that particular pool and model set. They do not establish reliability across independently constructed fingerprints or a broader population of unrelated models.

  3. I additionally noticed substantial degradation under routine deployment changes. E.g. 4-bit quantization reduces TMR from 98% to 21%.

Given these unresolved concerns, I believe the paper, in its current form, falls substantially short of the standard required for acceptance at ICLR. I therefore maintain my rejection recommendation (2: weak rejection, leaning towards 1) if no additional justifications. Please prioritize these additional experiments.

Reviewer_49TkOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper delves into the model copyright protection from the perspective of the hallucination-based fingerprinting. The authors propose the framework named SinkMark, which leverages attention-sink-induced hallucinations to verify model ownership. Extensive experiments on various LVLMs and their variants demonstrate the effectiveness of SinkMark.

Critical strengths
  1. Interesting perspective and clear motivation: This paper explores model copyright protection from the perspective of hallucination, turning an undesirable model behavior into a useful signal for copyright verification.

  2. Concise method design: SinkMark integrates three main modules, including Cross-Layer Sink, Target Semantic Binding, and Target Induction. Moreover, this method does not need permanent modifications to the protected model's parameters, which is practical in real-world scenarios.

  3. Novel experiment settings: The authors conduct extensive experiments under various model modifications, including LoRA fine-tuning, full fine-tuning, pruning, and quantization. These evaluations provide empirical evidence supporting the robustness of SinkMark.

Critical weaknesses

See Questions

Questions for authors
  1. I am very curious about the robustness of SinkMark against intentional fingerprint evasion strategies. Could the authors provide further analysis or experiments involving input preprocessing or output rewriting?

  2. Could the authors further analyze why SinkMark experiences substantial performance degradation under certain model modifications, such as full fine-tuning on TextVQA and 4-bit quantization of Qwen2.5-VL-7B?

  3. Could the authors further evaluate SinkMark on commercial closed-source LVLMs, such as GPT, Gemini, and Claude, to assess its fingerprint specificity and practical applicability in real-world black-box settings?

1 comment
AuthorsOpenReview ↗

Response to Reviewer : Evasion, Performance Degradation, and Closed-Source Evaluation

Thank you for the constructive comments.

Q-1. I am very curious about the robustness of SinkMark against intentional fingerprint evasion strategies. Could the authors provide further analysis or experiments involving input preprocessing or output rewriting?

RESPONSE: Appendix A.12 adds image preprocessing and DIPPER rewriting on LLaVA-1.5-7B and eight derivatives. Each model uses the same 200 images paired with five targets (1,000 image-target pairs). Probes and target assignments remain fixed, without re-optimization.

JPEG (quality 86), bicubic resizing (336 to 332 to 336), and Gaussian blur (radius 0.55) reduce source TMR from 98% to 53%, 46%, and 59% (Table 8). Derivative ranges are 21-49%, 14-19%, and 17-33%, respectively. These operations substantially weaken matching, especially resizing.

DIPPER rewrites every response, including initially unmatched responses. Lexical/order diversity settings 20/20, 40/40, 60/60, and 80/80 yield source TMRs of 96%, 92%, 87%, and 75%, versus 98% before rewriting (Table 10). Derivative TMRs range from 11% to 69%. Denominators remain the original fixed evaluation sets.

The tables below reproduce the TMR values; percentage-point changes are omitted for space. These non-adaptive operations are tested separately. We do not claim robustness to combined or target-aware attacks. A lost whole-word match may reflect paraphrasing rather than removal of the target concept (Section 6).

Image preprocessing (manuscript Table 8). L/F denote LoRA/full fine-tuning; Gaussian denotes blur.

ModelDefaultResizeJPEGGaussian
LLaVA-1.5-7B98%46%53%59%
L-MathV52%19%33%31%
L-PaintingForm62%14%42%33%
L-TextVQA33%14%28%17%
L-V7W57%18%49%30%
F-MathV48%15%27%23%
F-PaintingForm71%17%29%33%
F-TextVQA32%16%21%18%
F-V7W57%18%27%31%

DIPPER rewriting (manuscript Table 10). Column headings denote lexical/order diversity; baseline means no rewriting.

ModelBaseline20/2040/4060/6080/80
LLaVA-1.5-7B98%96%92%87%75%
L-MathV52%47%29%41%23%
L-PaintingForm62%52%51%48%43%
L-TextVQA33%26%21%25%20%
L-V7W57%54%54%52%42%
F-MathV48%48%41%34%14%
F-PaintingForm71%67%69%66%63%
F-TextVQA32%14%29%18%11%
F-V7W57%53%50%50%47%

Q-2. Could the authors further analyze why SinkMark experiences substantial performance degradation under certain model modifications, such as full fine-tuning on TextVQA and 4-bit quantization of Qwen2.5-VL-7B?

RESPONSE: Section 5.3 now distinguishes observed degradation from possible explanations. On TextVQA, Qwen TMR falls from 73% under LoRA to 20% under full fine-tuning (53 percentage points), versus 17-34 points on its other tasks (Table 3). LLaVA instead has similarly low LoRA/full TMRs (33%/32%). Persistence therefore depends on the task/model combination, not uniformly on full fine-tuning. Text-grounded adaptation may change response preferences or weaken probe-target associations; we have not isolated the cause.

Qwen TMR falls from 98% unquantized to 52% at 8-bit and 21% at 4-bit (Figure 4); LLaVA retains 80%/76% from a 98% baseline. Quantization may alter target-token likelihoods, attention, or generated prefixes, but these are hypotheses. Without controlled before/after measurements, we cannot establish a causal mechanism. We explicitly acknowledge these task/model-dependent and quantization limitations.

Q-3. Could the authors further evaluate SinkMark on commercial closed-source LVLMs, such as GPT, Gemini, and Claude, to assess its fingerprint specificity and practical applicability in real-world black-box settings?

RESPONSE: Appendix A.10 adds GPT-5-mini, Gemini-3.8-flash, and Claude 3.5 Sonnet as unrelated commercial negative controls. Each receives the same 200 fixed image-suffix probes through its black-box interface, without re-optimization. Case-insensitive, whole-word matching for the single expression "microscope" yields 0/200 matches on each model (0% empirical TMR).

This supports low activation on these models and probes, not universal zero FPR or generalization to other target-expression sets. Construction still requires white-box source access; verification uses returned responses only. Because these commercial models are unrelated controls, the experiment does not demonstrate API-only construction or positive ownership identification of a commercial derivative (Appendix A.10; Section 6).

We again thank the reviewer for the helpful comments.

Reviewer_Qv4NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduce SinkMark, an fingerprinting method for LVLM that induces target hallucination through cross-layer attention sinks. SinkMark jointly optimizes both image perturbation and suffixes tokens. The results show that SinkMark provides reliable and model-specific target hallucinations and remains robust to fine-tuning, pruning, and quantization.

Critical strengths
  • The problem setting is interesting

  • The performance is strong in the evaluated setup with 98% TMR on source models and 0% on evaluated unrelated models on both LLaVA and Qwen families.

  • Extensive experiments also covers fine-tuning, pruning, and quantization robustnesses.

Critical weaknesses
  • Limited novelty: Targeted hallucinations as fingerprinting are studied in prior work (PLA, AGDI). Attention-sink-induced hallucinations were explored in Mirage.

  • The paper lacks a controlled comparison with standard PGD+GCG+min–max optimization without sink-specific objectives making the incremental benefit unclear.

  • The theoretical analysis establishes only a generic local attention-margin stability condition, not robustness of target hallucinations. Also, the sink selection implementation is based on hidden-state similarity rather than attention concentration.

  • Performance degrades considerably in some settings (21% TMR on Qwen under 4-bit quantization). Robustness to common deployment changes, such as image preprocessing, system prompt template variations, and decoding configurations is not clear.

Questions for authors
  • Can we have an ablation study to see the actual contribution of sink-specific objectives in the overall framework?

  • How do image preprocessing, prompt-template variations, and decoding configurations affect the performance of SinkMark?

2 comments
AuthorsOpenReview ↗

Response to Reviewer (Part 1/2): Novelty, Ablations, and Method Clarifications

We thank the reviewer for the constructive comments and suggestions.

Part 1 addresses Q-1 to Q-3; Part 2 addresses Q-4 to Q-6.

Q-1. Limited novelty: Targeted hallucinations as fingerprinting are studied in prior work (PLA, AGDI). Attention-sink-induced hallucinations were explored in Mirage.

RESPONSE: We agree that neither targeted-response fingerprinting nor attention-sink-induced hallucination is individually new. SinkMark's contribution is to explicitly incorporate hallucination-associated attention dynamics into private behavioral fingerprint construction.

Distinction from PLA and AGDI. PLA optimizes adversarial trigger images for predefined responses; AGDI additionally uses image-text semantic alignment. Although unsupported target responses may be hallucinations, these methods do not explicitly optimize hallucination-associated sink behavior. SinkMark selects a potential aggregation position using cross-layer hidden-state similarity, encourages attention concentration and agreement among cross-layer candidate distributions, and associates the selected representation with a private target through semantic binding and post-sink induction (Section 4.1). Similarity is a selection heuristic, not a direct measurement of an existing sink.

Distinction from Mirage. Mirage studies sink-based hallucination attacks and cross-model transferability. SinkMark seeks ownership verification: persistence in protected-model derivatives and low activation on unrelated models. It combines sink-specific objectives with constrained image-text search and robustness-oriented parameter updates.

Evidence and revisions. Table 5 supports the combined design beyond generic PGD+GCG+min-max optimization (see Q-2). Manuscript Table 1 reports 98% source TMR on both primary models and no observed matches on evaluated unrelated models. These findings do not establish that attention concentration alone causes the gains or guarantees lineage specificity. We clarify the distinctions in the Introduction/Related Work and add the controlled ablation in Section 5.4.

Q-2. The paper lacks a controlled comparison with standard PGD+GCG+min–max optimization without sink-specific objectives, making the incremental benefit unclear.

RESPONSE: Section 5.4/Table 5 now compares generic probe optimization (GPO) with the complete sink-guided design (GPO+SSO). Both retain PGD image updates, constrained GCG suffix search, and temporary adversarial parameter updates. GPO uses sink-independent, teacher-forced target-expression cross-entropy. It omits attention concentration, cross-layer candidate consistency, target binding, predecessor guidance, and no-suffix attention control. GPO+SSO uses the complete objective, including post-sink target induction.

Adding the sink-guided design improves derivative TMR by 21, 21, 20, and 29 percentage points. The component ablations also reduce TMR. These comparisons support the combined design; because multiple objectives and guidance procedures differ, they do not isolate attention concentration's causal contribution.

Table 1. Derivative TMR on fully fine-tuned Qwen2.5-VL-7B (manuscript Table 5). GPO uses PGD, GCG, and min-max updates; SSO denotes sink-specific objectives.

MethodsMathVPaintingFormTextVQAV7W
GPO14%18%0%28%
GPO+SSO35%39%20%57%
GPO+SSO (w/o Cross-Layer Sink)15%13%5%31%
GPO+SSO (w/o Target Binding)17%13%0%15%

Q-3. The theoretical analysis establishes only a generic local attention-margin stability condition, not robustness of target hallucinations. Also, the sink selection implementation is based on hidden-state similarity rather than attention concentration.

RESPONSE: We agree with both observations. Section 3.1 now presents a local, conditional result: with a fixed decoding context and bounded attention-logit changes, Eq. 3 preserves the dominant position if its margin exceeds twice the perturbation bound. This generic attention-margin property does not prove target-hallucination persistence. Model modifications may violate the bound or change the decoding prefix; Section 5.3 assesses robustness empirically.

Section 4.1 separates candidate selection from attention shaping. Equation 8 averages hidden-state similarity across selected layers, following Mirage, to choose a potential aggregation position; it neither measures attention nor assumes an existing strong sink. Appendix A.1 converts similarity scores into candidate distributions (Eq. 13), aligns them through Jensen-Shannon consistency (Eq. 14), and encourages attention to the candidate (Eq. 15). Target binding and post-sink induction are separate objectives (Eqs. 9-10). Figure 16 is a qualitative visualization, not causal validation. Section 6 acknowledges that sink induction may fail on other attention architectures.

AuthorsOpenReview ↗

Response to Reviewer (Part 2/2): Robustness and Deployment Variations

This continues Part 1 and addresses Q-4 to Q-6.

Q-4. Performance degrades considerably in some settings (21% TMR on Qwen under 4-bit quantization). Robustness to common deployment changes, such as image preprocessing, system prompt template variations, and decoding configurations is not clear.

RESPONSE: We agree that the Qwen2.5-VL-7B result under 4-bit quantization is a substantial limitation. Figure 4 reports 21% TMR at 4-bit and 52% at 8-bit for Qwen, compared with 76% and 80% for LLaVA. Section 5.3 discusses possible changes in target-token likelihoods or attention behavior without claiming that these mechanisms have been verified.

We additionally evaluate deployment variations with fixed probes in Appendix A.12. On the LLaVA source model, JPEG compression, resizing, and Gaussian blur reduce TMR from 98% to 53%, 46%, and 59% (Table 8); two role-based system prompts yield 90% and 88% (Table 9). The temperature/top-pp sweep yields 32–84% TMR (Figure 18). Thus, the experiments identify meaningful preprocessing and decoding sensitivity rather than uniform deployment robustness. Section 6 acknowledges these limitations; our detailed protocol and derivative results are given in Q-6 below.

Q-5. Can we have an ablation study to see the actual contribution of sink-specific objectives in the overall framework?

RESPONSE: Section 5.4/Table 5 adds a sink-independent GPO control and component ablations. GPO+SSO achieves 35%, 39%, 20%, and 57% TMR, versus 14%, 18%, 0%, and 28% for GPO. Removing cross-layer modeling gives 15%, 13%, 5%, and 31%; removing target binding gives 17%, 13%, 0%, and 15%. This supports the combined objective and component groups, without isolating every attention or guidance term. Table 1 in Part 1 (Q-2) provides the full comparison.

Q-6. How do image preprocessing, prompt-template variations, and decoding configurations affect the performance of SinkMark?

RESPONSE: Appendix A.12 evaluates deployment variations using fixed probes without re-optimization. The image-preprocessing and system-prompt experiments use 200 images paired with five targets (1,000 image–target pairs), applying the same case-insensitive, whole-word matching against each target's accepted expressions before and after modification. The decoding sweep is separately reported on 200 fixed image probes in Appendix A.12.4; its denominator differs from the 1,000 image–target pairs in the preprocessing and system-prompt evaluations.

Image preprocessing (A.12.1, Table 8). JPEG compression (quality 86, 4:4:4 sampling), bicubic resizing (336 to 332 to 336), and Gaussian blur (radius 0.55) yield source-model TMRs of 53%, 46%, and 59%, respectively, compared with 98% under the default inputs. Across the eight derivatives, the corresponding ranges are 21–49%, 14–19%, and 17–33%, demonstrating sensitivity to image transformations.

System prompts (A.12.2, Table 9). Clinical-psychologist and technical-interviewer prompts yield source TMRs of 90% and 88%. Across the two prompts, TMR ranges from 23–52% for LoRA derivatives and 30–58% for fully fine-tuned derivatives.

Decoding (A.12.4, Figure 18). Sweeping temperature and top-p over the values 0.2, 0.4, 0.6, 0.8, and 1.0 yields TMRs of 32–84%; the result is 32% when both temperature and top-p equal 1.0. These tested configurations show substantial variation, without establishing robustness to arbitrary or adaptively chosen deployment settings.

Table 2 and Figure 1 reproduce the corresponding results.

Table 2. Deployment changes on LLaVA and its derivatives (TMR; entries reproduced from manuscript Tables 8–9). Blur denotes Gaussian blur. Clinical and Technical denote the two role-based system prompts. Preprocessing and prompt results use 1,000 image–target pairs per model with fixed target-specific matching rules. Each modification is evaluated separately.

ModelDefaultResizeJPEGBlurClinicalTechnical
LLaVA-1.5-7B98%46%53%59%90%88%
LoRA fine-tuning
MathV52%19%33%31%42%41%
PaintingForm62%14%42%33%48%46%
TextVQA33%14%28%17%23%27%
V7W57%18%49%30%43%52%
Full fine-tuning
MathV48%15%27%23%33%47%
PaintingForm71%17%29%33%54%58%
TextVQA32%16%21%18%31%30%
V7W57%18%27%31%44%39%

Figure 1. Temperature/top-pp sweep (reproduced from manuscript Figure 18). Probes and the target-matching rule are fixed across decoding configurations.

Figure location: Revised manuscript, Appendix A.12.4, Figure 18, PDF page 25.

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

Related papers

Discussion (0)

Sign in to comment.