acceptodds
Under review as a conference paper at ICLR 2027

REVEAL: Fingerprinting to Detect Model Reuse in Black-Box Ensembles

Abstract

Unauthorized reuse of a proprietary deep neural network (DNN) or its derived variants can be concealed within an ensemble, where prediction aggregation obscures individual model behavior. Existing fingerprinting methods primarily target standalone models, and their ownership-indicative signals can become substantially less distinguishable after aggregation. To address this challenge, we propose REVEAL, a model fingerprinting framework for detecting model REuse in black-box ensembles Via fingerprints generated through Ensemble-Aware Learning. Specifically, REVEAL derives fixed verification targets from naturally occurring prediction errors of the protected model and optimizes fingerprint queries without modifying model parameters. Its key idea is to reinforce the protected model's target responses while eliciting aggregation-neutral responses from independently trained reference models, which serve as surrogates for unrelated ensemble members. For single-label classification, reference predictions are driven toward uniformity to limit competing class preferences. For multi-label classification, reference probabilities for the selected target labels are driven slightly below the decision threshold, allowing a strong protected-model response to influence the final label decision. Ownership verification requires only the ensemble's returned labels and aggregates matches with the expected fingerprint targets into an ownership score. Extensive experiments demonstrate that REVEAL substantially mitigates the verification degradation caused by ensemble aggregation and effectively detects reuse of both protected models and their derived variants.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Model Fingerprinting, Model Ownership Verification, Model Ensembles

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 4 reviews, above 27% of the others.
Reviewer_qG9MOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the detection of unauthorized model reuse in black-box ensembles and proposes the REVEAL framework. The method optimizes fingerprint queries to reinforce the target responses of the protected model while reducing the interference of unrelated ensemble members in the aggregation process, thereby enabling ownership verification using only the labels returned by the ensemble. The paper considers both single-label and multi-label classification and empirically demonstrates the effectiveness of REVEAL in ensemble settings compared with conventional fingerprinting methods designed for standalone models. Overall, the topic is practically relevant and the methodological rationale is relatively clear. However, the effectiveness of the proposed approach still depends substantially on the representativeness of the reference models, the assumption of equal-weight probability averaging, and the generalization of the fingerprints to derived models. These issues require further justification.

Critical strengths

(1) Existing model-fingerprinting research has focused primarily on independently deployed standalone models, whereas this paper considers the more challenging scenario in which a protected model is concealed within a black-box ensemble. This extends the scope of research on model ownership verification.

(2) REVEAL develops different reference-model objectives for single-label and multi-label classification according to their respective aggregation mechanisms. Moreover, it performs verification without modifying model parameters and relies only on the labels returned by the ensemble, demonstrating desirable non-intrusiveness and methodological completeness.

Critical weaknesses

(1) The aggregation-neutral responses induced on the available reference models may not transfer to unknown ensemble members in a black-box ensemble. The authors should clarify how the reference models are obtained, state the assumptions required for their use, and evaluate the generalization of the learned fingerprints across different architectures, datasets, and training strategies.

(2) If the reference models assign excessively high probabilities to the target label, an ensemble composed entirely of unrelated models may output that label, increasing false positives. Conversely, if these probabilities are too low, they may overly dilute the protected model’s response, increasing false negatives. The authors should characterize the feasible operating conditions and conduct sensitivity analyses with respect to the decision threshold, target probability, ensemble size, and model-weight distribution.

(3) The optimization simultaneously requires strong responses from the protected model and near-threshold neutral responses from the reference models, and these objectives may conflict. The authors should report the per-query optimization success rate, the distribution of the resulting probability margins relative to the decision threshold, the optimization failure rate, and the stability of the results across different reference-model subsets and random seeds.

(4) The fingerprint queries are optimized only with respect to the original protected model, whereas fine-tuning, pruning, and knowledge distillation may substantially alter the model’s fingerprint responses. The authors should explain why the learned fingerprint signals are expected to survive such modifications and systematically evaluate detection performance under different modification types and modification strengths.

(5) The current framework is primarily developed for equally weighted ensembles. It remains unclear whether REVEAL remains effective under non-uniform model weights or alternative aggregation rules. The authors should analyze the minimum detectable weight of the protected model and its relationship to ensemble size, weight distribution, and fingerprint-signal strength.

(6) Under matched query budgets and false-positive rates, the authors should compare REVEAL with existing non-intrusive fingerprinting methods and backdoor watermarking approaches. They should also report the computational cost of fingerprint generation to demonstrate that the performance gains from aggregation-aware optimization justify its additional overhead.

Questions for authors

(1) How are the reference models obtained, and why can they represent unknown members of a black-box ensemble? In the multi-label setting, is it feasible to reliably control the target probability to just below the decision threshold?

(2) Is REVEAL applicable only to equal-weight probability averaging? How do ensemble size, non-uniform weights, and the protected model’s weight affect detection performance, and what is the minimum detectable weight wmin⁡w_{\min}?

(3) Why should fingerprints optimized on the original model transfer to variants produced by fine-tuning, pruning, and knowledge distillation?

2 comments
AuthorsOpenReview ↗

Response (1/2)

We thank the reviewer for the constructive comments and suggestions. We appreciate the opportunity to clarify the assumptions, design rationale, generalization capability, and practical applicability of REVEAL.

1. Response to Concern 1 and Question 1: Reference-model construction and generalization

As described in Sec. 4.1 and Appendix C, REVEAL uses independently trained ResNet-18 models with different random initializations as references. These models are trained on the same task and dataset as the protected model, without accessing its parameters or outputs, and are used exclusively during fingerprint generation.

Importantly, REVEAL does not require the reference models to share architectures with the actual ensemble members. In our evaluation, unrelated ensemble members are independently and uniformly sampled from nine architectures spanning ResNet, MobileNet, EfficientNet, and DenseNet families, with all evaluation models excluded from fingerprint generation. The strong AUC results across CIFAR-100, PET, VOC, and COCO (Tables 1–3) demonstrate effective verification on ensembles containing unseen models with diverse architectures. Moreover, Appendix B.1 shows that increasing the reference pool from 1 to 4 improves AUC from 0.891 to 0.916 on PET and from 0.872 to 0.891 on VOC, supporting the benefit of reference-model diversity.

While exact aggregation neutrality for arbitrary unseen models is not required or guaranteed, these results establish the effectiveness of reference-based fingerprint generation in the evaluated heterogeneous ensembles. We will further examine unseen models' response distributions and clarify the method's applicability conditions in the revised version.

2. Response to Concern 2 and Question 1: Feasible operating conditions and reference-model target probabilities

We first clarify that reference models are used exclusively during fingerprint generation and are excluded from all testing ensembles. Therefore, their target-label probabilities do not directly determine the predictions of independent ensembles.

For multi-label classification, the target probability is intentionally set slightly below the prediction threshold (ρr=0.45<τ=0.5\rho_r = 0.45 < \tau = 0.5), aiming to reduce false ownership matches while preserving sufficient target-label evidence when the protected model is present. We agree that this balance depends on the ensemble composition and aggregation weights.

Importantly, our existing experiments already investigate several relevant factors. Appendix B.4 (Fig. 8) evaluates sensitivity to the target probability, while Appendix B.5 (Fig. 9) examines ensemble sizes from 2 to 10. Furthermore, our additional unequal-weight experiments show that REVEAL maintains AUCs of 0.972 on CIFAR-100 and 0.980 on PET even when the protected model's coefficient is only 0.1. These results provide empirical evidence of REVEAL's effectiveness across the evaluated operating conditions. We will clarify these findings and further examine threshold sensitivity in the revised version.

To further assess practical verification reliability, we conducted an ownership-threshold sensitivity analysis over τo∈0.1,0.2,…,0.9\tau_o \in \\{0.1, 0.2, \ldots, 0.9\\}, which will be included in the revised version. At the default threshold τo=0.5\tau_o = 0.5, REVEAL achieves zero FPR on both CIFAR-100 and PET, with FNRs of 0.020 and 0.189, respectively. Notably, CIFAR-100 achieves zero FPR at τo=0.3\tau_o = 0.3, whereas PET requires τo=0.5\tau_o = 0.5, illustrating the dataset-dependent trade-off between false positives and false negatives. Together, these results characterize REVEAL's effectiveness across the evaluated operating conditions. We will incorporate these findings and clarify the corresponding applicability conditions in the revised manuscript.

Table R1. FPR / FNR under different ownership-decision thresholds.

ThresholdCIFAR-100PET
0.10.208 / 0.0000.568 / 0.069
0.20.064 / 0.0030.204 / 0.143
0.30.000 / 0.0090.028 / 0.180
0.40.000 / 0.0140.008 / 0.186
0.50.000 / 0.0200.000 / 0.189
0.60.000 / 0.0630.000 / 0.226
0.70.000 / 0.1600.000 / 0.309
0.80.000 / 0.4400.000 / 0.540
0.90.000 / 0.9290.000 / 0.940

3. Response to Concern 3: Feasibility and stability of joint fingerprint optimization

We would like to clarify that the protected-model and reference-model objectives are designed to encourage distinct responses, rather than requiring every fingerprint to exactly reach the prescribed target probabilities. Although these objectives may compete during optimization, REVEAL jointly optimizes them to enhance the separability between pirated and independent ensembles. The strong verification AUCs reported in Tables 1–3 provide empirical evidence that this joint optimization is effective.

AuthorsOpenReview ↗

Response (2/2)

4. Response to Concern 4 and Question 3: Generalization to modified models

Regarding fingerprint robustness under model modifications, although fingerprints are generated using the original protected model, common modifications such as FT, PR, NFT, and PFT may preserve part of its learned decision behavior. By reinforcing the protected model's target-specific responses, REVEAL aims to generate stronger ownership signals that can remain distinguishable after such modifications, rather than relying on fragile responses near decision boundaries.

5. Response to Concern 5 and Question 2: Unequal model weights and minimum detectable weight

Impact of non-uniform weight strategy. To further investigate the impact of non-uniform model weights, we conducted additional experiments on CIFAR-100 and PET using five-member ensembles. We varied the reused model's coefficient among 0.1, 0.2, and 0.3. REVEAL consistently achieved the highest AUC across all evaluated settings. Even at the smallest tested coefficient of 0.1, REVEAL achieved AUCs of 0.972 and 0.980 on CIFAR-100 and PET, respectively.

Table R2. Ownership-verification AUC under weighted probability averaging with different coefficients assigned to the reused protected model.

DatasetwP\boldsymbol{w_P}AKHIPGuardIrisFPLiteGuardREVEAL
CIFAR-1000.10.807 ± 0.0540.681 ± 0.0730.715 ± 0.0790.903 ± 0.0360.972 ± 0.008
CIFAR-1000.20.866 ± 0.0330.739 ± 0.0230.787 ± 0.0660.934 ± 0.0160.988 ± 0.004
CIFAR-1000.30.879 ± 0.0300.778 ± 0.0270.812 ± 0.0580.943 ± 0.0171.000 ± 0.000
PET0.10.787 ± 0.0340.621 ± 0.0820.834 ± 0.0960.856 ± 0.0420.980 ± 0.002
PET0.20.832 ± 0.0240.683 ± 0.0510.865 ± 0.0900.887 ± 0.0380.995 ± 0.003
PET0.30.851 ± 0.0220.713 ± 0.0280.876 ± 0.0650.915 ± 0.0191.000 ± 0.000

Regarding the minimum detectable weight, our experimental results show that REVEAL maintains strong verification performance even at the smallest evaluated aggregation coefficient of 0.1, achieving AUCs of 0.972 and 0.980 on CIFAR-100 and PET, respectively. As expected, AUC gradually decreases as the reused model's weight decreases, reflecting the reduced contribution of its ownership evidence to the aggregated predictions. However, a universal minimum detectable weight is difficult to define, as it depends on the desired verification performance and the specific ensemble configuration.

From a practical perspective, an adversary who incorporates a stolen model into an ensemble typically aims to exploit its predictive capabilities. Assigning an extremely small weight to the stolen model would reduce its influence on the ensemble's final predictions, potentially diminishing the benefit of reusing it. Therefore, reducing the stolen model's contribution to evade ownership verification may introduce a trade-off between verification evasion and the practical utility of model reuse.

6. Response to Concern 6: Baseline comparisons and computational cost

We would like to clarify that REVEAL has already been compared with four existing non-intrusive fingerprinting methods (IPGuard, AKH, IrisFP, and LiteGuard) under the same fingerprint query budget of 100 (Sec. 4.1). The results in Tables 1–3 demonstrate REVEAL's superior verification performance under ensemble reuse.

Regarding backdoor watermarking, such methods typically require modifying the protected model to embed ownership information, whereas REVEAL performs non-intrusive fingerprinting without modifying the protected model. Therefore, these approaches operate under different ownership-verification assumptions and are not directly comparable under the same deployment constraints.

We agree that quantitative profiling is necessary to assess whether the additional optimization overhead is justified. As shown in Table R3, all methods incur relatively low fingerprint-generation costs, with REVEAL achieving competitive efficiency. Moreover, the online verification cost is primarily determined by the number of fingerprint queries. Since all methods use the same query budget, they incur comparable black-box query costs during ownership verification.

Table R3. Fingerprint generation time for different methods under the same query budget (N=100N = 100).

MethodGeneration time
IPGuard1 m 20 s
AKH35 s
IrisFP1 m 31 s
LiteGuard1 m 16 s
REVEAL55 s
Reviewer_PkSwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes REVEAL, a model fingerprinting method for detecting reuse of a protected model or its derived variants within black-box ensembles. It optimizes fingerprint queries by reinforcing the protected model’s target responses while shaping reference-model responses toward neutral outputs, aiming to preserve fingerprint signals under ensemble aggregation. The method requires no modification of the protected model and uses only predicted labels for ownership verification. Experiments on four datasets covering single-label and multi-label classification evaluate its effectiveness under equal-weight probability averaging, including heterogeneous ensembles and several model modifications. The results show improved discrimination over the evaluated baselines in different settings.

Critical strengths
  1. The paper addresses a meaningful, underexplored problem with a well-motivated design. Detecting protected-model reuse within black-box ensembles is practically relevant and challenging. The proposed combination of protected-model response reinforcement and reference-model response shaping is well motivated.

  2. The empirical results support the method’s effectiveness within the evaluated setting. Across four datasets, REVEAL demonstrates the ability to distinguish pirated ensembles from independent ensembles under equal-weight probability averaging. Further experiments demonstrate its ability to identify protected-model reuse in heterogeneous ensembles and retain substantial discriminative power under several modifications, including fine-tuning and pruning.

Critical weaknesses
  1. False-positive and false-negative control remains unverified, limiting confidence in ownership verification. AUC measures the overall separability of pirated and independent ensembles, but does not establish the reliability of ownership decisions at a predetermined threshold. The paper sets the ownership-decision threshold to 0.50.5 without sufficiently justifying this choice or reporting the corresponding false-positive rate (FPR) and false-negative rate (FNR) on the test ensembles. Consequently, the current results do not establish whether the method can reliably control false ownership claims against independent ensembles while maintaining adequate detection performance.
  2. The method has a mechanistic dependence on probability averaging. The paper restricts its scope to equal-weight probability averaging and does not evaluate hard voting, another common aggregation rule. Hard voting converts the reference models’ “neutral” responses into definite votes. In particular, under per-label majority voting for multi-label classification, reference responses below the decision threshold become votes against the target label, potentially suppressing the fingerprint signal. Whether changing the aggregation rule can evade the proposed ownership verification while preserving task utility therefore remains an important unresolved risk.
Questions for authors
  1. What FPR and FNR does the method achieve on the test ensembles at the fixed ownership-decision threshold?

  2. Although Figure 6 shows some benefit from reference-model response shaping, the improvement appears relatively modest compared with protected-model response reinforcement. Could the authors provide further evidence that this component is particularly important for mitigating fingerprint dilution as the ensemble size increases?

  3. Does replacing probability averaging with hard voting substantially affect verification performance?

  4. Does the method remain effective when the protected model uses an architecture other than ResNet-18?

2 comments
AuthorsOpenReview ↗

Response (1/2)

We thank the reviewer for the constructive comments. We appreciate the opportunity to clarify the scope of REVEAL, the interpretation of its evaluation metrics, and the role of reference-model response shaping.

1. Response to Concern 1 and Question 1: False-positive/false-negative rates and ownership-decision threshold

Our primary objective is to generate fingerprints that distinguish pirated ensembles from independent ensembles based on their ownership scores. We therefore adopt AUC as the primary metric, as it measures this discriminability across all possible decision thresholds, enabling threshold-independent comparisons across fingerprinting methods.

We use τo=0.5\tau_o=0.5 as the default operating threshold, but acknowledge that this choice may not provide the optimal balance between FPR and FNR across different settings. In practice, the threshold can be calibrated using an independent validation set to achieve the desired false-positive control. To further address this concern, we will report FPR and FNR at different ownership thresholds, together with the ownership-score distributions of pirated and independent ensembles, to evaluate practical verification reliability.

2. Response to Concern 2 and Question 3: Probability averaging versus hard voting

First, regarding majority voting, we evaluated REVEAL following the two ensemble settings in Sec. 4.2. In the first setting, each pirated ensemble directly incorporates the original protected model alongside independently trained, unrelated members, whereas each independent ensemble consists entirely of unrelated models. As shown in Table R1, REVEAL achieves AUCs of 1.000, 0.995, 0.978, and 1.000 on CIFAR-100, PET, VOC, and COCO, respectively, outperforming all applicable baselines.

Table R1. Ownership-verification AUC under majority voting when an ensemble directly reuses the original protected model.

MethodCIFAR-100PETVOCCOCO
IPGuard0.809 ± 0.0370.695 ± 0.046----
AKH0.857 ± 0.0420.882 ± 0.028----
IrisFP0.878 ± 0.0440.919 ± 0.089----
LiteGuard0.961 ± 0.0180.900 ± 0.0150.904 ± 0.0460.917 ± 0.060
REVEAL1.000 ± 0.0000.995 ± 0.0110.978 ± 0.0191.000 ± 0.000

In the second setting, we evaluated ensembles containing modified versions of the protected model, obtained through FT, PR, KD, NFT, and PFT. As reported in Table R2, REVEAL achieves AUCs ranging from 0.928 to 1.000 across all 16 dataset–modification combinations involving FT, PR, NFT, and PFT, outperforming all applicable baselines. KD remains particularly challenging on PET, VOC, and COCO, where neither REVEAL nor existing methods consistently achieves strong verification performance. One possible reason is that KD may alter the protected model's distinctive response patterns, making ownership evidence harder to preserve after ensemble aggregation.

Table R2. Ownership-verification AUC under majority voting when an ensemble reuses a modified version of the protected model.

DatasetMethodFTPRKDNFTPFT
CIFAR-100IPGuard0.447 ± 0.1080.765 ± 0.1230.773 ± 0.0580.471 ± 0.2520.410 ± 0.119
CIFAR-100AKH0.707 ± 0.0580.991 ± 0.0380.795 ± 0.0500.562 ± 0.1120.742 ± 0.134
CIFAR-100IrisFP0.964 ± 0.0030.478 ± 0.0960.443 ± 0.0820.951 ± 0.0310.900 ± 0.005
CIFAR-100LiteGuard0.913 ± 0.0160.957 ± 0.0070.809 ± 0.0290.879 ± 0.0230.929 ± 0.004
CIFAR-100REVEAL1.000 ± 0.0001.000 ± 0.0000.990 ± 0.0110.996 ± 0.0051.000 ± 0.000
PETIPGuard0.423 ± 0.0810.441 ± 0.1020.459 ± 0.1250.346 ± 0.0640.230 ± 0.157
PETAKH0.773 ± 0.1430.609 ± 0.1320.291 ± 0.0590.720 ± 0.0520.824 ± 0.076
PETIrisFP0.715 ± 0.0680.671 ± 0.1270.647 ± 0.1230.769 ± 0.0730.711 ± 0.165
PETLiteGuard0.851 ± 0.0280.938 ± 0.0130.326 ± 0.0560.884 ± 0.0210.927 ± 0.012
PETREVEAL0.928 ± 0.0060.991 ± 0.0030.495 ± 0.0400.998 ± 0.0011.000 ± 0.000
VOCLiteGuard0.867 ± 0.0530.792 ± 0.0430.317 ± 0.0590.928 ± 0.0380.841 ± 0.031
VOCREVEAL0.967 ± 0.0290.971 ± 0.0240.249 ± 0.0690.978 ± 0.0160.985 ± 0.018
COCOLiteGuard0.867 ± 0.0240.938 ± 0.0110.441 ± 0.0870.851 ± 0.0290.901 ± 0.041
COCOREVEAL0.957 ± 0.0120.992 ± 0.0060.479 ± 0.0780.976 ± 0.0160.974 ± 0.018
AuthorsOpenReview ↗

Response (2/2)

3. Response to Question 2: Contribution of reference-model response shaping as ensemble size increases

Protected-model reinforcement and reference-model response shaping serve complementary roles: the former strengthens ownership evidence, while the latter reduces interference from unrelated ensemble members. As shown in Fig. 6, reference-model shaping provides an additional AUC improvement of approximately 0.04–0.05 on PET and VOC. Meanwhile, Fig. 9 shows that increasing the ensemble size from 2 to 10 reduces AUC from 0.936 to 0.858 on PET and from 0.939 to 0.835 on VOC, highlighting the challenge of fingerprint dilution. Since larger ensembles introduce more competing predictions, reference-model shaping is designed to mitigate this interference. To further quantify its contribution, we will evaluate REVEAL with and without reference-model shaping across different ensemble sizes under identical settings.

4. Response to Question 4: Generalization to different protected-model architectures

From a methodological perspective, REVEAL does not rely on architecture-specific layers or internal structures. Fingerprint generation uses the protected model's output probabilities and their gradients with respect to the input queries, while ownership verification requires only the suspect ensemble's returned labels. Thus, the formulation can be applied to other differentiable classification architectures without changing its fundamental objectives. As shown in Table R3, REVEAL maintains strong verification performance under the MobileNet-v2 architecture, consistently outperforming all baselines and further demonstrating its effectiveness across different model architectures.

Table R3. REVEAL's verification performance under standalone-model and ensemble scenarios with the MobileNetV2 architecture of the protected model.

MethodScenarioCIFAR-100PETVOCCOCO
IPGuardStandalone1.0001.000----
IPGuardEnsemble0.7890.707----
AKHStandalone1.0001.000----
AKHEnsemble0.8870.824----
IrisFPStandalone1.0001.000----
IrisFPEnsemble0.8270.886----
LiteGuardStandalone1.0001.0001.0001.000
LiteGuardEnsemble0.9540.9270.9360.908
REVEALStandalone1.0001.0001.0001.000
REVEALEnsemble1.0001.0000.9901.000
Reviewer_ytQCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces REVEAL, a method for detecting unauthorized reuse of a protected model within black-box ensembles. It creates fingerprint queries from inputs the protected model misclassifies and refines them to strengthen its target predictions while reducing interference from independent models. The method does not modify model parameters and verifies reuse using only the ensemble’s returned labels. Its main contribution is an ensemble-aware fingerprinting approach with a common verification rule for single-label and multi-label tasks.

Experiments on CIFAR-100, PET, VOC, and COCO show near-perfect detection of direct reuse, with AUCs of 0.994–1.000 using 100 queries. Results are also strong for several fine-tuning and pruning variants. However, detection of distilled derivatives is substantially weaker, with AUCs of 0.464, 0.231, and 0.465 on PET, VOC, and COCO, respectively. These findings support REVEAL’s effectiveness for direct reuse but show limited robustness to distillation.

Critical strengths
  1. REVEAL assigns complementary objectives to the protected model and unrelated references. Uniform reference probabilities avoid competing-class preference in single-label averaging, while in multi-label prediction reference probabilities slightly below the decision threshold keep unrelated ensembles negative yet allow a strong protected-model response to affect the final decision. These task-specific constructions follow the respective decision rules and enable label-only verification.

  2. The experiments cover four datasets spanning different task settings, four mainstream baseline methods, and five common model modification approaches, together with comprehensive ablation studies that systematically verify the necessity of each component and the impact of hyperparameters.

Critical weaknesses
  1. REVEAL is formulated and evaluated exclusively for ensembles that combine predictive probability vectors through equal-weight averaging. Its reference-response shaping is specifically designed for this aggregation rule, and its effectiveness under alternative strategies, such as unequal weighting, majority voting, or stacking, remains unexamined. In particular, responses that are approximately neutral under probability averaging need not remain neutral under voting or a learned aggregation function. Furthermore, fingerprint generation uses independently trained ResNet-18 reference models. Although the paper evaluates ensembles drawn from nine architectures and explicitly excludes the reference models from these ensembles, it does not fully specify how architectural combinations are sampled or report performance across different architectural compositions.

  2. The evaluation considers standard model modifications, including fine-tuning and noise injection followed by fine-tuning, but does not assess an adaptive adversary who knows REVEAL’s design and deliberately selects or optimizes modifications to suppress fingerprint evidence while preserving predictive utility. Robustness to the evaluated modifications therefore does not establish resistance to such targeted attack.

Questions for authors
  1. Can REVEAL be extended beyond equal-weight probability averaging to other aggregation methods such as majority voting or unequal-weight averaging? What changes to fingerprint generation or reference-response shaping would be required, and is there preliminary empirical or theoretical evidence supporting these extensions?

  2. How are architectures sampled and combined from the nine-architecture pool? Please clarify the sampling distribution and whether repeated architectures are allowed within an ensemble. Does verification performance vary with architectural composition, and are any corresponding results available?

  3. What justifies setting the ownership-decision threshold to τo=0.5\tau_o = 0.5? Was this threshold calibrated on independent validation data, and what false-positive and true-positive rates does it achieve across datasets and model modifications?

3 comments
AuthorsOpenReview ↗

Response (1/3)

We thank the reviewer for the constructive comments and suggestions. We appreciate the opportunity to clarify REVEAL's aggregation assumptions, architectural diversity, robustness against adaptive attacks, and ownership-decision threshold.

1. Response to Concern 1 & Question 1: Generalization to alternative aggregation strategies

We thank the reviewer for raising this important question. We first clarify that the original formulation of REVEAL focuses on equal-weight probability averaging, as defined in Sec. 2. To further investigate its applicability beyond equal-weight averaging, we conducted additional experiments under majority voting and unequal-weight probability averaging.

First, regarding majority voting, we evaluated REVEAL following the two ensemble settings in Sec. 4.2. In the first setting, each pirated ensemble directly incorporates the original protected model alongside independently trained, unrelated members, whereas each independent ensemble consists entirely of unrelated models. As shown in the following table, REVEAL achieves AUCs of 1.000, 0.995, 0.978, and 1.000 on CIFAR-100, PET, VOC, and COCO, respectively, outperforming all applicable baselines.

Table R1. Ownership-verification AUC under majority voting when an ensemble directly reuses the original protected model.

MethodCIFAR-100PETVOCCOCO
IPGuard0.809 ± 0.0370.695 ± 0.046----
AKH0.857 ± 0.0420.882 ± 0.028----
IrisFP0.878 ± 0.0440.919 ± 0.089----
LiteGuard0.961 ± 0.0180.900 ± 0.0150.904 ± 0.0460.917 ± 0.060
REVEAL1.000 ± 0.0000.995 ± 0.0110.978 ± 0.0191.000 ± 0.000

In the second setting, we evaluated ensembles containing modified versions of the protected model, obtained through FT, PR, KD, NFT, and PFT. As reported in Table R2, REVEAL achieves AUCs ranging from 0.928 to 1.000 across all 16 dataset–modification combinations involving FT, PR, NFT, and PFT, outperforming all applicable baselines. KD remains particularly challenging on PET, VOC, and COCO, where neither REVEAL nor existing methods consistently achieves strong verification performance. One possible reason is that KD may alter the protected model's distinctive response patterns, making ownership evidence harder to preserve after ensemble aggregation.

Table R2. Ownership-verification AUC under majority voting when an ensemble reuses a modified version of the protected model.

DatasetMethodFTPRKDNFTPFT
CIFAR-100IPGuard0.447 ± 0.1080.765 ± 0.1230.773 ± 0.0580.471 ± 0.2520.410 ± 0.119
CIFAR-100AKH0.707 ± 0.0580.991 ± 0.0380.795 ± 0.0500.562 ± 0.1120.742 ± 0.134
CIFAR-100IrisFP0.964 ± 0.0030.478 ± 0.0960.443 ± 0.0820.951 ± 0.0310.900 ± 0.005
CIFAR-100LiteGuard0.913 ± 0.0160.957 ± 0.0070.809 ± 0.0290.879 ± 0.0230.929 ± 0.004
CIFAR-100REVEAL1.000 ± 0.0001.000 ± 0.0000.990 ± 0.0110.996 ± 0.0051.000 ± 0.000
PETIPGuard0.423 ± 0.0810.441 ± 0.1020.459 ± 0.1250.346 ± 0.0640.230 ± 0.157
PETAKH0.773 ± 0.1430.609 ± 0.1320.291 ± 0.0590.720 ± 0.0520.824 ± 0.076
PETIrisFP0.715 ± 0.0680.671 ± 0.1270.647 ± 0.1230.769 ± 0.0730.711 ± 0.165
PETLiteGuard0.851 ± 0.0280.938 ± 0.0130.326 ± 0.0560.884 ± 0.0210.927 ± 0.012
PETREVEAL0.928 ± 0.0060.991 ± 0.0030.495 ± 0.0400.998 ± 0.0011.000 ± 0.000
VOCLiteGuard0.867 ± 0.0530.792 ± 0.0430.317 ± 0.0590.928 ± 0.0380.841 ± 0.031
VOCREVEAL0.967 ± 0.0290.971 ± 0.0240.249 ± 0.0690.978 ± 0.0160.985 ± 0.018
COCOLiteGuard0.867 ± 0.0240.938 ± 0.0110.441 ± 0.0870.851 ± 0.0290.901 ± 0.041
COCOREVEAL0.957 ± 0.0120.992 ± 0.0060.479 ± 0.0780.976 ± 0.0160.974 ± 0.018
AuthorsOpenReview ↗

Response (2/3)

1. Response to Concern 1 & Question 1 (continued): Unequal-weight probability averaging

Second, regarding unequal-weight probability averaging, we conducted additional experiments on CIFAR-100 and PET using ensembles containing five members. The aggregation coefficient of the reused protected model was varied among 0.1, 0.2, and 0.3. As shown in Table R3, REVEAL consistently achieves the highest AUC across both datasets and all evaluated weight settings. Even at the smallest evaluated coefficient of 0.1, REVEAL achieves AUCs of 0.972 and 0.980 on CIFAR-100 and PET, respectively, compared with 0.903 and 0.856 for the strongest baseline, LiteGuard.

Table R3. Ownership-verification AUC under weighted probability averaging with different coefficients assigned to the reused protected model.

DatasetwPw_PAKHIPGuardIrisFPLiteGuardREVEAL
CIFAR-1000.10.807 ± 0.0540.681 ± 0.0730.715 ± 0.0790.903 ± 0.0360.972 ± 0.008
CIFAR-1000.20.866 ± 0.0330.739 ± 0.0230.787 ± 0.0660.934 ± 0.0160.988 ± 0.004
CIFAR-1000.30.879 ± 0.0300.778 ± 0.0270.812 ± 0.0580.943 ± 0.0171.000 ± 0.000
PET0.10.787 ± 0.0340.621 ± 0.0820.834 ± 0.0960.856 ± 0.0420.980 ± 0.002
PET0.20.832 ± 0.0240.683 ± 0.0510.865 ± 0.0900.887 ± 0.0380.995 ± 0.003
PET0.30.851 ± 0.0220.713 ± 0.0280.876 ± 0.0650.915 ± 0.0191.000 ± 0.000

The effectiveness under unequal-weight averaging is also consistent with the design of REVEAL. For single-label classification, a uniform reference-model response contributes equally to all classes regardless of its aggregation weight, thereby introducing no competing class preference. Under majority voting, however, probabilistic responses are converted into discrete votes, so the same neutrality property is not theoretically guaranteed. Nevertheless, our additional experiments demonstrate that REVEAL retains strong verification performance under the evaluated majority-voting settings. These results provide empirical evidence that REVEAL's verification effectiveness is not limited to equal-weight probability averaging. We will incorporate these results and clarify the applicability of the current fingerprint-generation design to different aggregation mechanisms in the revised version. Stacking, which involves a learned aggregation function, remains outside the scope of the current evaluation.

2. Concern 1 & Question 2: Architectural sampling and composition

Architectural sampling and composition.

We would like to clarify that our evaluation considers randomly constructed ensembles rather than a predefined set of architectural combinations.

Specifically, for each testing ensemble, we randomly select the ensemble size KK from 2,3,4,5\\{2,3,4,5\\}. The architectures of independently trained ensemble members are uniformly and independently sampled from the nine architectures listed in Sec. 4.1. Repeated architectures are allowed within an ensemble, but models sharing the same architecture are trained independently with different random initializations. For pirated ensembles, one member is the protected model or its derived variant, while the remaining K−1K-1 unrelated members are sampled using this procedure. Independent ensembles consist entirely of independently trained, unrelated models.

This random construction allows us to evaluate REVEAL across diverse architectural combinations without manually selecting particular combinations that might favor our method. In addition, all testing models are independent of the reference models used during fingerprint generation. Therefore, the reported AUCs reflect REVEAL's verification performance across randomly sampled ensembles containing models from different architectural families, rather than performance on a small set of predefined architecture combinations.

We will clarify the sampling procedure in the revised version.

AuthorsOpenReview ↗

Response (3/3)

3. Concern 2: Robustness against adaptive model modifications

We acknowledge that an attacker with knowledge of REVEAL's fingerprint generation mechanism could potentially construct proxy fingerprints and modify the stolen model to evade ownership verification. However, under our threat model, the owner-generated fingerprints remain confidential and are therefore unavailable to the attacker. Although the attacker may independently generate proxy fingerprints, these queries do not necessarily coincide with the owner's fingerprints. Consequently, modifications optimized to evade verification using proxy fingerprints may not generalize to the owner's confidential fingerprints. Furthermore, suppressing the model behaviors exploited by REVEAL may also affect legitimate predictions, introducing a potential trade-off between evasion effectiveness and predictive utility. A systematic evaluation of such adaptive attacks is beyond the scope of our current study.

4. Question 3: Ownership-decision threshold and detection performance

We thank the reviewer for raising this question.

Our primary objective is to generate fingerprints that distinguish pirated ensembles from independent ensembles based on their ownership scores (Sec. 3.3). Thus, we adopt AUC as the primary metric, as it evaluates this discriminability across all possible decision thresholds. In contrast, TPR and FPR depend on the selected threshold and thus reflect performance at a specific operating point.

To further address the reviewer's concern, we conducted a threshold-sensitivity analysis by considering ownership thresholds ranging from 0.1 to 0.9. Table R4 reports the FPR and FNR at different ownership decision thresholds. At the default threshold τo=0.5\tau_o=0.5, REVEAL achieves zero FPR on both CIFAR-100 and PET, with FNRs of 0.020 and 0.189, respectively. Notably, CIFAR-100 already achieves zero FPR at τo=0.3\tau_o=0.3, whereas PET requires τo=0.5\tau_o=0.5, illustrating dataset-dependent operating characteristics. Increasing the threshold reduces false positives but increases missed detections. These results support τo=0.5\tau_o=0.5 as a conservative default across the evaluated datasets, while application-specific thresholds can be calibrated using an independent validation set to achieve the desired FPR--FNR trade-off.

Table R4. FPR / FNR under different ownership-decision thresholds.

ThresholdCIFAR-100PET
0.10.208 / 0.0000.568 / 0.069
0.20.064 / 0.0030.204 / 0.143
0.30.000 / 0.0090.028 / 0.180
0.40.000 / 0.0140.008 / 0.186
0.50.000 / 0.0200.000 / 0.189
0.60.000 / 0.0630.000 / 0.226
0.70.000 / 0.1600.000 / 0.309
0.80.000 / 0.4400.000 / 0.540
0.90.000 / 0.9290.000 / 0.940
Reviewer_9kWgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to detect whether a protected model, or one of its derived variants (e.g., obtained via fine-tuning, pruning, or knowledge distillation), has been incorporated without authorization into a black-box ensemble. In this setting, the ensemble averages member outputs, which can weaken the ownership signals captured by existing fingerprinting methods. The authors propose REVEAL, which consists of two steps.

  • Initialization: samples naturally misclassified by the protected model serve as fingerprint queries, and the erroneous predictions are retained as target labels.
  • Query optimization: the queries are optimized to reinforce the protected model’s response to the targets. At the same time, a set of independently trained reference models, used as surrogates for unknown ensemble members, is driven toward “aggregation-neutral” outputs: near-uniform distributions in the single-label setting, and target-label probabilities slightly below the decision threshold in the multi-label setting.

Verification requires only the labels returned by the ensemble, and the fraction of matched targets serves as the ownership score. Experiments on CIFAR-100, PET, VOC, and COCO show that REVEAL clearly outperforms existing fingerprinting methods in the ensemble setting and remains robust to most model modifications.

Critical strengths
  1. The problem is practically meaningful and well motivated. The paper empirically demonstrates the performance degradation of existing fingerprinting methods under ensemble deployment, which shows that this scenario merits dedicated study.
  2. The method is simple and intuitive. The idea of keeping unrelated members “neutral” under aggregation is easy to follow, and the objectives are tailored to the different decision mechanisms of single-label and multi-label classification.
  3. Coverage of multi-label classification. Most existing fingerprinting work focuses on single-label tasks, so this extension adds value.
  4. Strong results under most model modifications. Under fine-tuning, pruning, and their combinations, the verification performance in the ensemble setting generally matches or exceeds the strongest baseline.
  5. The evaluation protocol is well designed. Members of the test ensembles span nine architectures and are strictly separated from the reference models, and the appendix provides a fairly comprehensive analysis of the main hyperparameters.
Critical weaknesses
  1. Unconstrained optimization and no adaptive-attack evaluation. The paper does not appear to specify any perturbation constraint on the fingerprint queries, so the optimized queries may drift away from the natural data distribution. In addition, since the adversary has access to the outputs of all ensemble members, it may be relatively easy for them to notice the anomalous pattern in which one member is highly confident while the others are close to uniform, and then exclude that member.
  2. Only equal-weight probability averaging is considered. The aggregation rule is chosen by the adversary, who could instead use, for example, majority voting or weighted averaging. The paper neither discusses this limitation nor evaluates such alternatives.
  3. Aggregation neutrality is not directly verified, and the evidence for generalization is limited. The paper reports only AUC. It does not show how close the outputs of the optimized reference models, or those of the members in the test ensembles, actually are to the uniform distribution. In addition, all reference models used during fingerprint generation are ResNet-18, so it is not clear whether the neutral responses generalize to members with other architectures.
  4. The source of the performance gain is unclear. In the ablation study (Fig. 6), removing the reference-model term reduces the AUC by only about 0.04, which suggests that most of the gain comes from reinforcing the protected model’s responses. Moreover, in the single-label setting there is no comparison with a generic reference-model regularization (for example, simply lowering the reference models’ probabilities on the target class). It is therefore hard to tell whether the uniform-distribution design is truly necessary.
Questions for authors
  1. If any part of my understanding above is inaccurate, I would appreciate it if the authors could point it out.
  2. Is any perturbation constraint (e.g., an ε-norm bound or a pixel-range constraint) imposed when optimizing the fingerprint queries? If an adversary identifies fingerprint queries by exploiting output discrepancies among ensemble members, would REVEAL remain effective?
  3. If the adversary adopts a different aggregation rule, such as majority voting or weighted averaging, would the mechanism of REVEAL still hold?
  4. Since all reference models used during fingerprint generation are ResNet-18, how do the authors confirm that members with other architectures in the test ensembles also produce approximately uniform responses?
  5. Removing the reference-model term reduces the AUC by only about 0.04. How do the authors view the actual contribution of the aggregation-neutral design?
4 comments
AuthorsOpenReview ↗

Response (1/3)

We thank the reviewer for the constructive comments. We appreciate the opportunity to clarify the scope of REVEAL, the rationale behind its aggregation-neutral design, and the limitations of our current evaluation.

1. Response to Concern 1 and Question 2: Perturbation constraints and adaptive attacks

We thank the reviewer for raising these important concerns.

First, regarding perturbation constraints, we agree with the reviewer that constraining the perturbation magnitude is important, as some fingerprints may be out of the distribution. Thus, we will incorporate a perturbation regularization term into the fingerprint-generation objective to limit the deviation of optimized queries from their original inputs in the revised version. We also would like to emphasize that our work primarily focuses on an underexplored research problem, i.e., detecting unauthorized model reuse by trying to preserve ownership evidence under ensemble aggregation. To the best of our knowledge, REVEAL is the first fingerprinting framework specifically designed to address this problem.

Second, regarding adaptive attacks, such a strategy would require knowledge of REVEAL's fingerprint generation mechanism. However, prediction discrepancies among ensemble members are not exclusive to fingerprint queries. They may also arise from clean data samples, particularly in heterogeneous ensembles. Therefore, identifying and filtering fingerprint queries solely based on such discrepancies could lead to false positives, inadvertently filtering clean inputs and potentially degrading the ensemble's predictive performance.

2. Response to Concern 2 and Question 3: Generalization to other aggregation rules.

We thank the reviewer for raising this important concern. As specified in Sec. 2, REVEAL was originally designed and evaluated for equal-weight probability averaging. We agree that the response-shaping objectives do not automatically guarantee the same behavior under different aggregation rules. Thus, we conducted additional experiments to examine REVEAL's effectiveness under majority voting and unequal-weight probability averaging settings.

Majority voting.

We evaluated REVEAL under majority voting using the two ensemble evaluation settings described in Sec. 4.2. First, we considered direct reuse of the original protected model: each pirated ensemble contains the protected model alongside independently trained, unrelated members, whereas each independent ensemble consists entirely of independently trained, unrelated models. As shown in Table R1, REVEAL achieves AUCs of 1.000, 0.995, 0.978, and 1.000 on CIFAR-100, PET, VOC, and COCO, respectively, outperforming all applicable baselines.

Table R1. Ownership-verification AUC under majority voting when an ensemble directly reuses the original protected model.

MethodCIFAR-100PETVOCCOCO
IPGuard0.809 ± 0.0370.695 ± 0.046----
AKH0.857 ± 0.0420.882 ± 0.028----
IrisFP0.878 ± 0.0440.919 ± 0.089----
LiteGuard0.961 ± 0.0180.900 ± 0.0150.904 ± 0.0460.917 ± 0.060
REVEAL1.000 ± 0.0000.995 ± 0.0110.978 ± 0.0191.000 ± 0.000
AuthorsOpenReview ↗

# Response (2/3)

2. Concern 2 and Question 3 (continued)

Second, we evaluated robustness to model modifications under majority voting. For each modification type (FT, PR, KD, NFT, or PFT), the corresponding derived model was incorporated into an ensemble alongside independently trained, unrelated members, as in the Ensemble setting of Sec. 4.2. The resulting ensembles were distinguished from independent ensembles using their ownership scores. As shown in Table R2, REVEAL achieves AUCs from 0.928 to 1.000 across all 16 dataset–modification combinations involving FT, PR, NFT, and PFT, outperforming all applicable baselines in these settings.

Table R2. Ownership-verification AUC under majority voting when an ensemble reuses a modified version of the protected model.

DatasetMethodFTPRKDNFTPFT
CIFAR-100IPGuard0.447 ± 0.1080.765 ± 0.1230.773 ± 0.0580.471 ± 0.2520.410 ± 0.119
CIFAR-100AKH0.707 ± 0.0580.991 ± 0.0380.795 ± 0.0500.562 ± 0.1120.742 ± 0.134
CIFAR-100IrisFP0.964 ± 0.0030.478 ± 0.0960.443 ± 0.0820.951 ± 0.0310.900 ± 0.005
CIFAR-100LiteGuard0.913 ± 0.0160.957 ± 0.0070.809 ± 0.0290.879 ± 0.0230.929 ± 0.004
CIFAR-100REVEAL1.000 ± 0.0001.000 ± 0.0000.990 ± 0.0110.996 ± 0.0051.000 ± 0.000
PETIPGuard0.423 ± 0.0810.441 ± 0.1020.459 ± 0.1250.346 ± 0.0640.230 ± 0.157
PETAKH0.773 ± 0.1430.609 ± 0.1320.291 ± 0.0590.720 ± 0.0520.824 ± 0.076
PETIrisFP0.715 ± 0.0680.671 ± 0.1270.647 ± 0.1230.769 ± 0.0730.711 ± 0.165
PETLiteGuard0.851 ± 0.0280.938 ± 0.0130.326 ± 0.0560.884 ± 0.0210.927 ± 0.012
PETREVEAL0.928 ± 0.0060.991 ± 0.0030.495 ± 0.0400.998 ± 0.0011.000 ± 0.000
VOCLiteGuard0.867 ± 0.0530.792 ± 0.0430.317 ± 0.0590.928 ± 0.0380.841 ± 0.031
VOCREVEAL0.967 ± 0.0290.971 ± 0.0240.249 ± 0.0690.978 ± 0.0160.985 ± 0.018
COCOLiteGuard0.867 ± 0.0240.938 ± 0.0110.441 ± 0.0870.851 ± 0.0290.901 ± 0.041
COCOREVEAL0.957 ± 0.0120.992 ± 0.0060.479 ± 0.0780.976 ± 0.0160.974 ± 0.018

These results provide empirical evidence that REVEAL retains strong ownership-score discrimination under majority voting for direct reuse and the four modifications tested here. Although hard voting converts probability vectors into discrete votes, the proposed method remains effective in these evaluated settings.

Unequal-weight probability averaging.

We further evaluated REVEAL under unequal-weight probability averaging on CIFAR-100 and PET. Each ensemble contains five models, and the coefficient of the reused protected model is set to 0.10.1, 0.20.2, or 0.30.3. 0.20.2 corresponds to equal weighting, while 0.10.1 assigns the reused model a smaller coefficient than any unrelated member. Table R3 reports the verification AUCs.

Table R3. Ownership-verification AUC under weighted probability averaging with different aggregation coefficients assigned to the reused protected model.

DatasetwPw_PAKHIPGuardIrisFPLiteGuardREVEAL
CIFAR-1000.10.807 ± 0.0540.681 ± 0.0730.715 ± 0.0790.903 ± 0.0360.972 ± 0.008
CIFAR-1000.20.866 ± 0.0330.739 ± 0.0230.787 ± 0.0660.934 ± 0.0160.988 ± 0.004
CIFAR-1000.30.879 ± 0.0300.778 ± 0.0270.812 ± 0.0580.943 ± 0.0171.000 ± 0.000
PET0.10.787 ± 0.0340.621 ± 0.0820.834 ± 0.0960.856 ± 0.0420.980 ± 0.002
PET0.20.832 ± 0.0240.683 ± 0.0510.865 ± 0.0900.887 ± 0.0380.995 ± 0.003
PET0.30.851 ± 0.0220.713 ± 0.0280.876 ± 0.0650.915 ± 0.0191.000 ± 0.000

As shown, REVEAL achieves the highest AUC on both datasets at every tested coefficient. Even when the reused model's coefficient is only 0.10.1, REVEAL achieves AUCs of 0.9720.972 on CIFAR-100 and 0.9800.980 on PET, compared with 0.9030.903 and 0.8560.856 for LiteGuard, the strongest baseline in these settings. Increasing the coefficient to 0.30.3 raises REVEAL's AUC to 1.0001.000 on both datasets. These results demonstrate that REVEAL can preserve strong ownership-score discrimination even when the reused model is given a lower coefficient than each unrelated member. Moreover, since an adversary typically reuses a stolen model to benefit from its predictive capabilities, assigning it a negligibly small weight could undermine the utility of the model reuse itself, creating a potential trade-off between evading ownership verification and preserving the benefits of the stolen model.

AuthorsOpenReview ↗

Response (3/3)

3. Response to Concern 3 and Question 4: Aggregation neutrality and cross-architecture generalization

We would like to clarify that encouraging aggregation-neutral responses is one of the optimization terms for reducing interference from independently trained models during probability averaging when generating fingerprints. Since there are infinitely many independently trained models, we introduce a number of reference models when generating fingerprints. However, REVEAL does not require every unseen ensemble member to produce an exactly uniform distribution. What ultimately matters is whether the generated fingerprints preserve sufficient ownership evidence to distinguish pirated and independent ensembles.

To evaluate this capability, as described in Sec. 4.1, the architectures of unrelated members in the testing ensembles are independently and uniformly sampled from nine architectures, including ResNet, MobileNet, EfficientNet, and DenseNet families. These testing models are entirely excluded from fingerprint generation. Table 1 reports that REVEAL achieves AUCs of 1.000, 1.000, 0.994, and 1.000 on CIFAR-100, PET, VOC, and COCO, respectively, showing REVEAL's strong ownership-score discriminability across ensembles containing unseen models with diverse architectural compositions.

Regarding the adoption of AUC as a metric, our primary objective is to generate fingerprints that distinguish pirated ensembles from independent ensembles based on their ownership scores. Thus, as in existing fingerprinting approaches, we adopt AUC as the primary metric, as it measures this discriminability across all possible decision thresholds, enabling threshold-independent comparisons across fingerprinting methods. To address the reviewer's concern, we additionally report verification accuracy under different ownership-decision thresholds. The verification accuracy is defined as Accuracy=TP+TNTP+FP+TN+FN\mathrm{Accuracy} = \frac{TP + TN}{TP + FP + TN + FN},

where TPTP, FPFP, TNTN, and FNFN denote the numbers of true positives, false positives, true negatives, and false negatives, respectively.

As shown in Table R4, REVEAL maintains high verification accuracy across a relatively wide range of ownership-decision thresholds, indicating that its performance is not highly sensitive to threshold selection. In particular, it achieves accuracies above 0.96 on CIFAR-100 and 0.86 on PET for thresholds ranging from 0.3 to 0.6.

Table R4. Ownership-verification accuracy under different ownership-decision thresholds.

ThresholdCIFAR-100PET
0.10.9130.723
0.20.9720.832
0.30.9950.883
0.40.9920.888
0.50.9880.890
0.60.9630.868
0.70.9070.820
0.80.7430.685
0.90.4580.452

4. Response to Concern 4 and Question 5: Contribution of reference-model response shaping

We agree with the reviewer that protected-model reinforcement contributes more substantially to REVEAL's verification performance. This difference is consistent with their distinct roles: the protected-model term establishes the primary ownership evidence by reinforcing the protected model's target-specific responses, while the reference-model term further reduces interference from unrelated models during ensemble aggregation. Therefore, the reference-model term is intended to complement, rather than replace, protected-model reinforcement.

Regarding the uniform-distribution objective, its motivation is to reduce the influence of reference models on the aggregated prediction across all classes, rather than merely suppress their probabilities on the fingerprint target class. Under probability averaging, an exactly uniform prediction contributes the same constant to every class and therefore does not favor any competing class. In contrast, simply suppressing the target-class probability may redistribute probability mass toward other classes, potentially introducing stronger competition with the protected model's target prediction. This provides the theoretical motivation for adopting uniform-distribution shaping. We will further evaluate this design by comparing it with a target-class probability suppression baseline in the revised version.

We thank the authors for their response. Though the additional experimental results partially address our previous concerns, several important issues remain unresolved. For non-uniform ensemble aggregation, the robustness of the proposed method may be substantially weakened under unequal-weight averaging. Moreover, in majority-voting scenarios, the contribution of the protected model may be significantly weakened, particularly when the ensemble size exceeds (K=2). In addition, for the architectural composition of the ensemble, the architectures of the models used during the fingerprint generation phase should be explicitly specified. Overall, it provides valuable additional empirical evidence, but the remaining concerns lead me to maintain my current score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.