Pith. sign in

REVIEW 3 major objections 5 minor 54 references

HoneypotNet: Backdoor Attacks Against Model Extraction

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read Replacing a model's output layer with a honeypot layer lets defenders inject backdoors into stolen substitute models.

desk verdict Genuinely new 'attack as defense' with surprisingly strong hard-label results; the core UAP-transfer assumption is fragile and the paper undersells that fragility. read the letter →

arxiv 2501.01090 v1 pith:5DJD62XF submitted 2025-01-02 cs.CR cs.CV

classification cs.CRcs.CV
keywords modelextractionattacksbackdooradversarialdefenseuniversalperturbationbi-leveloptimizationownershipverificationattackashoneypotlayer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes HoneypotNet, a defense that attacks the attacker: instead of merely detecting model extraction, the model owner modifies the victim model's output layer so that any substitute model trained on those outputs inherits a backdoor. The backdoor is triggered by a universal adversarial perturbation, a small image perturbation that makes the substitute predict a fixed target class. The authors show on CIFAR10, CIFAR100, Caltech256, and CUBS200, across five extraction attacks, that attack success rates on substitute models range from 56.99% to 92.35% while clean accuracy stays close to the undefended model. The significance is a new 'attack as defense' paradigm: the defender gains both ownership verification and a reverse attack that can disrupt a stolen model.

What carries the argument

The load-bearing mechanism is the honeypot layer plus a universal adversarial perturbation (UAP) used as a trigger. The honeypot layer $H(x)=W\cdot F_{\text{feat}}(x)+b$ replaces the victim's classification head, so it controls only the outputs and adds few parameters. The trigger $\delta$ is found on a lightweight shadow model (ResNet18) and then reinforced by fine-tuning the honeypot layer so that the association 'input plus $\delta$ maps to target class' behaves like a normal function. The bi-level optimization formalizes this: the upper level asks the honeypot layer to keep clean accuracy while mapping triggered inputs to the target class, and the lower level trains the shadow model and updates the trigger. The paper's argument is that adversarial vulnerability transfers: a UAP optimized on the shadow model transfers through the poisonous probability vectors into any substitute model trained on them.

What would settle it

Train a substitute model from HoneypotNet-protected outputs using only hard labels and strong robustification, such as adversarial training or defensive distillation, on a dataset where the paper reports high attack success; if the attack success rate falls to the undefended baseline while clean accuracy stays high, the transferability claim fails. A simpler check is to repeat the KnockoffNets extraction with a vision-transformer substitute, where UAP transfer is typically weak.

Watch

Extended reading notes

Core claim

The central claim is that a backdoor can be injected into an extracted substitute model without ever poisoning the attacker's images or retraining the victim. The method replaces the victim's classification layer with a honeypot layer, a single fully connected layer that maps the victim's feature vector to a probability vector. A bi-level optimization loop alternates between training a shadow model on the honeypot layer's outputs, generating a universal adversarial perturbation that makes the shadow model predict the target class, and fine-tuning the honeypot layer to preserve normal accuracy while becoming sensitive to that perturbation. Because the substitute model is trained on the honeypot layer's poisoned probability vectors, it learns the perturbation as a normal input-output association and therefore reproduces the backdoor. The authors report that this works under soft-label extraction and also under the hard-label BlackBox Dissector, with verification accuracy and attack success rate both far above undefended baselines.

Load-bearing premise

The entire defense depends on the assumption that a universal adversarial perturbation computed on a small shadow model will transfer to the substitute model the attacker actually trains, using the poisoned probability vectors as the conduit.

Editorial extensions

If this is right

  • A model owner who deploys HoneypotNet can verify ownership by querying a suspect model with triggered inputs: if the suspect predicts the target class at an unusually high rate, it was likely extracted from the protected API.
  • The same trigger acts as a reverse attack key, forcing a stolen substitute model to misclassify any triggered input as the target class, which can deter or damage an attacker's deployment.
  • The defense requires no retraining of the victim and only a lightweight honeypot layer, so it can be applied to pre-trained models with low computational overhead.
  • The method remains effective under hard-label extraction, where the attacker trains on discrete labels rather than full probability vectors.
  • Trigger size is a tunable trade-off: larger triggers raise attack success but also begin to perturb the victim's own predictions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the trigger is a UAP, the scheme inherits the known fragility of adversarial transferability: an attacker who trains with a very different architecture (for instance a vision transformer) or uses adversarial training on the substitute may break the backdoor, a risk the paper acknowledges when reporting lower success on DenseNet121.
  • The 'attack as defense' framing suggests an escalation dynamic: extraction defenses that poison outputs could push attackers toward hard-label-only or output-filtering strategies, which in turn may make extraction harder or more expensive even when the backdoor fails.
  • The UAP-based trigger's resistance to detection and pruning suggests that backdoor defenses based on finding minimal norm patterns may need to be re-evaluated against triggers that align with the model's existing decision boundaries.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes HoneypotNet, a "attack as defense" technique against model extraction. The method replaces the victim model's classification layer with a honeypot layer and uses a shadow model plus a universal adversarial perturbation (UAP) trigger inside a bi-level optimization loop. The honeypot layer is fine-tuned to keep normal predictions while making triggered inputs map to a target class, and the paper claims that a substitute model trained on the honeypot layer's soft-label outputs inherits a backdoor. Experiments on CIFAR10, CIFAR100, CUBS200, and Caltech256 with five extraction attacks, including a hard-label attack, report verification accuracies from 52.29% to 92.61% and attack success rates from 56.99% to 92.35%, with additional tests on trigger size, substitute architecture, backdoor detection, and neuron pruning.

Significance. If the mechanism holds, HoneypotNet introduces a genuinely new defense paradigm: instead of only detecting or watermarking extraction, the defender uses the extraction process itself to implant a controllable backdoor, enabling both ownership verification and an active reverse attack. The paper's strengths are its broad experimental coverage (five extraction attacks, four datasets, multiple substitute architectures, hard-label setting) and its evaluation against two backdoor defenses (Cognitive Distillation and Reconstructive Neuron Pruning). The main weakness is that the load-bearing step, transfer of the UAP-triggered backdoor from the honeypot layer to a substitute trained only on clean pairs, is justified empirically through UAP transferability and shows noticeable sensitivity across architectures. The significance is therefore conditional: the idea is valuable, but the current evidence does not yet establish that the backdoor is reliably inherited beyond the specific training procedures and architecture ranges tested.

major comments (3)
  1. [Experimental Setup and Table 2] The claim that HoneypotNet preserves the victim model's utility is not directly supported by the main results. Table 2 reports Accc for the substitute models, not for the protected model after the classification layer is replaced by the honeypot layer. The original victim accuracies (91.56%, 71.57%, 77.11%, and 78.44%) are given, but the clean accuracy of the HoneypotNet-protected model is never tabulated. Please add the protected model's clean test accuracy for each dataset and trigger configuration so that the reader can verify that the honeypot layer does not degrade the served model.
  2. [Eq. (6), trigger masking] The mask operation in Eq. (6) appears inverted. If M is the binarized trigger mask that restricts the trigger to a specific location, then the triggered input should be (1-M)⊙x + M⊙δ, not M⊙x + (1-M)⊙δ as written. As printed, the equation places the trigger everywhere except the masked region. This needs to be corrected or the definition of M clarified, because the trigger generation step is central to the method's reproducibility.
  3. [Section 'Finetuning the Honeypot Layer', Eqs. (4)-(6), and Table 3] The backdoor transfer mechanism is the load-bearing step: a substitute trained only on clean pairs (x, H(x)) must inherit a trigger that was optimized on a shadow model and never appears in the substitute's training set. The paper asserts this via UAP transferability, but does not provide a controlled study of when transfer fails. Table 3 already shows architecture sensitivity (e.g., DenseNet121 reaches only 51.68% ASR on CIFAR10), and the shadow model and shadow dataset differ from the attacker's architecture and transfer set. Please add ablations over attacker-side training loss (e.g., KL divergence vs. cross-entropy), robust/adversarial training, shadow model architecture, and shadow model ensembles. These experiments would delimit the conditions under which the central claim holds and would separate the contribution of the BLO loop from simply training H to be adversarially vulnerable.
minor comments (5)
  1. [Algorithm 1, line 5] The loss function L' used in the extraction simulation step is never defined; please specify whether it is cross-entropy, KL divergence, or another objective.
  2. [Table 1] The symbols '!' and '%' used in Table 1 are not defined in the caption or in the surrounding text, which makes the comparison table hard to interpret.
  3. [Abstract and Introduction] The abstract says HoneypotNet protects against 'any malicious users' who attempt to extract the victim model, but the experiments cover only the standard soft-label and hard-label extraction pipelines. Please scope the claim to the tested threat model.
  4. [Figure 3] The curve labeled 'HoneypotNet' in Figure 3 is not defined in the caption; clarify whether it shows the protected model's clean accuracy after the honeypot layer is inserted.
  5. [Section 'Finetuning the Honeypot Layer'] The term 'universal adversarial perturbation' is used for a trigger optimized on a small verification set Dv and constrained by a fixed mask and location; this is not universal in the original sense of Moosavi-Dezfooli et al. and should be justified or renamed.

Circularity Check

0 steps flagged · score 0.0 of 10

HoneypotNet's backdoor transfer is an externally measured empirical result, not a prediction forced by construction; no load-bearing circularity found.

full rationale

The paper's central mechanism—a UAP δ optimized on a shadow model and a honeypot layer H fine-tuned so that H(x+δ) predicts the target class, with the backdoor expected to transfer to a substitute trained only on clean H(x)—is an empirical design, not a derivation that reduces to its own inputs. The BLO objective in Eqs. (4)-(6) optimizes H and δ on the defender's shadow set, but the reported ASR is measured on substitute models trained from scratch by five different extraction methods, including the hard-label BlackBox Dissector, and across four substitute architectures. The strong variation in results (e.g., DenseNet121 drops to 51.68% ASR on CIFAR10 while VGG16 reaches 97.16%) demonstrates that the evaluation is not a tautology: if the high ASR were forced by construction, architecture and extraction method would not matter. The transferability of the UAP is an explicit empirical assumption, acknowledged in the paper ('Given that our backdoor trigger is based on the transferability of UAP, its effectiveness can vary depending on the model architecture'), and its fragility is a robustness/correctness concern rather than a circularity. No load-bearing claim is justified only by a self-citation: the citations to Huang et al. (2023) and Li et al. (2023a), which share an author with this paper, are used solely as external evaluation tools (backdoor detection and neuron-pruning defenses), not to establish the central transfer mechanism. Thus no circular step is identified.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method introduces a honeypot layer and an optimized trigger, but these are computational artifacts rather than newly postulated physical entities. The free parameters listed above are hand-chosen or fitted values that the central claim depends on. The axioms are the threat model assumptions and the transferability assumptions that carry the mechanism; none are machine-checked or independently verified beyond the reported experiments.

free parameters (4)
  • Universal Adversarial Perturbation trigger delta = Not reported; optimized via Eq. (6)
    The trigger is learned on the shadow model and is the mechanism for backdoor injection; its exact values are dataset-dependent and not disclosed. The reported ASR depends entirely on this fitted perturbation.
  • Trigger mask size and location = 6x6 for CIFAR10/100; 28x28 for Caltech256/CUBS200; upper-left corner, 4px offset
    Chosen by hand; Figure 3 shows ASR varies with trigger size, so this choice materially affects the central claim.
  • Target class y_target = Last class of each dataset
    Chosen for simplicity; the specific class may affect transferability and verification accuracy.
  • Ownership verification threshold = 10%
    Used to declare extraction when Accv on the target class exceeds 10%; chosen by hand without sensitivity analysis.
assumptions (5)
  • domain assumption Adversarial perturbations transfer across architectures and training procedures (UAP transferability).
    Invoked in 'Finetuning the Honeypot Layer' and Eq. (6): the trigger is optimized on a ResNet18 shadow model and assumed to trigger substitute models of varied architectures and extraction methods. If transferability fails for some attacker, the backdoor is not inherited.
  • domain assumption The attacker trains a substitute model on soft-label output vectors H(x) using a standard loss such as cross-entropy or KL divergence.
    Eq. (1) defines the substitute as minimizing L(F_hat(x), H(x)); the backdoor transfer mechanism relies on this standard training. Hard-label attackers are only tested with one specific method, BlackBox Dissector.
  • domain assumption Only the output layer can be changed, and the victim's training data is inaccessible, so defensive fine-tuning must use a proxy shadow set from a different distribution, CC3M.
    Threat Model states 'Only the output F(x) can be modified' and 'the victim model may be pre-trained, with its training data inaccessible'. The method's reliance on a shadow distribution is an unverified assumption that the proxy transfer set is representative enough.
  • domain assumption Attacker queries are indistinguishable from legitimate queries, so no query filtering is used.
    Threat Model assumption (2); this justifies output poisoning rather than detection-based defense.
  • ad hoc to paper A universal adversarial perturbation can act as a poisoning-free backdoor trigger without explicit injection into training images.
    The authors justify this by analogy to adversarial vulnerability, but it is a design postulate specific to this method and is not independently established for extracted models outside the reported experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HoneypotNet: Backdoor Attacks Against Model Extraction." pith.science (2026). https://pith.science/paper/5DJD62XF

@misc{pith2026250101090,
  author       = {Pith},
  title        = {Pith review of: HoneypotNet: Backdoor Attacks Against Model Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5DJD62XF}},
  note         = {Machine review of arXiv:2501.01090}
}
read the original abstract

Model extraction attacks are one type of inference-time attacks that approximate the functionality and performance of a black-box victim model by launching a certain number of queries to the model and then leveraging the model's predictions to train a substitute model. These attacks pose severe security threats to production models and MLaaS platforms and could cause significant monetary losses to the model owners. A body of work has proposed to defend machine learning models against model extraction attacks, including both active defense methods that modify the model's outputs or increase the query overhead to avoid extraction and passive defense methods that detect malicious queries or leverage watermarks to perform post-verification. In this work, we introduce a new defense paradigm called attack as defense which modifies the model's output to be poisonous such that any malicious users that attempt to use the output to train a substitute model will be poisoned. To this end, we propose a novel lightweight backdoor attack method dubbed HoneypotNet that replaces the classification layer of the victim model with a honeypot layer and then fine-tunes the honeypot layer with a shadow model (to simulate model extraction) via bi-level optimization to modify its output to be poisonous while remaining the original performance. We empirically demonstrate on four commonly used benchmark datasets that HoneypotNet can inject backdoors into substitute models with a high success rate. The injected backdoor not only facilitates ownership verification but also disrupts the functionality of substitute models, serving as a significant deterrent to model extraction attacks.

Figures

Figures reproduced from arXiv: 2501.01090 by the authors.

Figure 1
Figure 1. An illustration of our HoneypotNet defense. leakage risks associated with MLaaS platforms due to model extraction attacks (Orekondy, Schiele, and Fritz 2019a; Pal et al. 2020; Yu et al. 2020; Zhou et al. 2020; Wang et al. 2022; Lin et al. 2023; Zhao et al. 2023; Karmakar and Basu 2023; Liu et al. 2024; Yuan et al. 2024). In a model ex￾traction attack, an attacker approximates a black-box vic￾tim model by training a … view at source ↗
Figure 2
Figure 2. Overview of our HoneypotNet method. It replaces the classification layer of the victim model with a honeypot layer [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The impact of trigger size on the victim model, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The L1 norm distributions of the detected backdoor patterns for clean and backdoor samples by the backdoor detection method Cognitive Distillation (CD) (Huang et al. 2023). Arc. CIFAR10 CIFAR100 CUBS200 Caltech256 ResNet34 59.35 85.71 78.31 79.13 ResNet18 59.17 69.55 7…
Figure 5
Figure 5. Figure 5: Robustness of HoneypotNet against Reconstruc [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 44 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Chen, H.; Fu, C.; Zhao, J.; and Koushanfar, F. 2021. Proflip: Targeted trojan attack with progressive bit flips. In ICCV

  4. [4]

    Chen, K.; Lou, X.; Xu, G.; Li, J.; and Zhang, T. 2022. Clean-image backdoor: Attacking multi-label models with poisoned labels only. In ICLR

  5. [5]

    Chen, X.; Liu, C.; Li, B.; Lu, K.; and Song, D. 2017. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526

  6. [6]

    A.; Lu, Y

    Dziedzic, A.; Kaleem, M. A.; Lu, Y. S.; and Papernot, N. 2022. Increasing the cost of model extraction with calibrated proof of work. In ICLR

  7. [7]

    Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. 2014. Generative adversarial nets. In NIPS

  8. [8]

    J.; Shlens, J.; and Szegedy, C

    Goodfellow, I. J.; Shlens, J.; and Szegedy, C. 2015. Explaining and harnessing adversarial examples. In ICLR

Show all 54 references
  1. [9]

    Griffin, G.; Holub, A.; and Perona, P. 2007. Caltech-256 object category dataset

  2. [10]

    Gu, T.; Dolan-Gavitt, B.; and Garg, S. 2017. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733

  3. [11]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In CVPR

  4. [12]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. In NeurIPS

  5. [13]

    M.; and Bailey, J

    Huang, H.; Ma, X.; Erfani, S. M.; and Bailey, J. 2023. Distilling Cognitive Backdoor Patterns within an Image. In ICLR

  6. [14]

    Jha, R.; Hayase, J.; and Oh, S. 2023. Label poisoning is all you need. NeurIPS

  7. [15]

    A.; Chandrasekaran, V.; and Papernot, N

    Jia, H.; Choquette-Choo, C. A.; Chandrasekaran, V.; and Papernot, N. 2021. Entangled watermarks as a defense against model extraction. In USENIX Security Symposium

  8. [16]

    Jindal, A.; Goyal, V.; Anand, S.; and Arora, C. 2024. Army of Thieves: Enhancing Black-Box Model Extraction via Ensemble based sample selection. In WACV

  9. [17]

    Juuti, M.; Szyller, S.; Marchal, S.; and Asokan, N. 2019. PRADA: protecting against DNN model stealing attacks. In IEEE EuroS&P

  10. [18]

    Kariyappa, S.; Prakash, A.; and Qureshi, M. 2021. MAZE: Data-Free Model Stealing Attack Using Zeroth-Order Gradient Estimation. In CVPR

  11. [19]

    Kariyappa, S.; and Qureshi, M. K. 2020. Defending Against Model Stealing Attacks with Adaptive Misinformation. In CVPR

  12. [20]

    Karmakar, P.; and Basu, D. 2023. Marich: A Query-efficient Distributionally Equivalent Model Extraction Attack using Public Data. arXiv preprint arXiv:2302.08466

  13. [21]

    Kesarwani, M.; Mukhoty, B.; Arya, V.; and Mehta, S. 2018. Model extraction warning in mlaas paradigm. In ACSAC

  14. [22]

    Krizhevsky, A.; Hinton, G.; et al. 2009. Learning multiple layers of features from tiny images

  15. [23]

    Kullback, S.; and Leibler, R. A. 1951. On information and sufficiency. The annals of mathematical statistics

  16. [24]

    Li, Y.; Lyu, X.; Ma, X.; Koren, N.; Lyu, L.; Li, B.; and Jiang, Y.-G. 2023 a . Reconstructive neuron pruning for backdoor defense. In ICML

  17. [25]

    Li, Y.; Zhu, L.; Jia, X.; Jiang, Y.; Xia, S.-T.; and Cao, X. 2022. Defending against model stealing via verifying embedded external features. In AAAI

  18. [26]

    Li, Y.; Zhu, M.; Yang, X.; Jiang, Y.; Wei, T.; and Xia, S.-T. 2023 b . Black-box dataset ownership verification via backdoor watermarking. IEEE TIFS

  19. [27]

    Lin, Z.; Xu, K.; Fang, C.; Zheng, H.; Ahmed Jaheezuddin, A.; and Shi, J. 2023. QUDA: Query-Limited Data-Free Model Extraction. In ACM Asia CCS

  20. [28]

    Liu, Y.; Chen, X.; Liu, C.; and Song, D. 2017. Delving into transferable adversarial examples and black-box attacks. In ICLR

  21. [29]

    Liu, Y.; Lee, W.-C.; Tao, G.; Ma, S.; Aafer, Y.; and Zhang, X. 2019. Abs: Scanning neural networks for back-doors by artificial brain stimulation. In ACM SIGSAC CCS

  22. [30]

    Liu, Y.; Wen, R.; Backes, M.; and Zhang, Y. 2024. Efficient Data-Free Model Stealing with Label Diversity. arXiv preprint arXiv:2404.00108

  23. [31]

    Lv, P.; Ma, H.; Chen, K.; Zhou, J.; Zhang, S.; Liang, R.; Zhu, S.; Li, P.; and Zhang, Y. 2024. MEA-Defender: A Robust Watermark against Model Extraction Attack. In IEEE S&P

  24. [32]

    Moosavi-Dezfooli, S.-M.; Fawzi, A.; Fawzi, O.; and Frossard, P. 2017. Universal adversarial perturbations. In CVPR

  25. [33]

    Orekondy, T.; Schiele, B.; and Fritz, M. 2019 a . Knockoff Nets: Stealing Functionality of Black-Box Models. In CVPR

  26. [34]

    Orekondy, T.; Schiele, B.; and Fritz, M. 2019 b . Prediction Poisoning: Towards Defenses Against DNN Model Stealing Attacks. In ICLR

  27. [35]

    Pal, S.; Gupta, Y.; Shukla, A.; Kanade, A.; Shevade, S.; and Ganapathy, V. 2020. ACTIVETHIEF: Model Extraction Using Active Learning and Unannotated Public Data. In AAAI

  28. [36]

    B.; and Swami, A

    Papernot, N.; McDaniel, P.; Goodfellow, I.; Jha, S.; Celik, Z. B.; and Swami, A. 2017. Practical black-box attacks against machine learning. In ACM AsiACCS

  29. [37]

    Ribeiro, M.; Grolinger, K.; and Capretz, M. A. 2015. Mlaas: Machine learning as a service. In IEEE ICMLA

  30. [38]

    Rong, D.; Shen, S.; Fu, X.; Qian, P.; Chen, J.; He, Q.; Fu, X.; and Wang, W. 2024. Clean-image Backdoor Attacks. arXiv preprint arXiv:2403.15010

  31. [39]

    Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; et al. 2015. Imagenet large scale visual recognition challenge. IJCV

  32. [40]

    Sanyal, S.; Addepalli, S.; and Babu, R. V. 2022. Towards data-free model stealing in a hard label setting. In CVPR

  33. [41]

    S.; and Goldstein, T

    Shafahi, A.; Najibi, M.; Xu, Z.; Dickerson, J.; Davis, L. S.; and Goldstein, T. 2020. Universal adversarial training. In AAAI

  34. [42]

    Sharma, P.; Ding, N.; Goodman, S.; and Soricut, R. 2018. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL

  35. [43]

    Szegedy, C.; Zaremba, W.; Sutskever, I.; Bruna, J.; Erhan, D.; Goodfellow, I.; and Fergus, R. 2014. Intriguing properties of neural networks. In ICLR

  36. [44]

    Tan, J.; Zhong, N.; Qian, Z.; Zhang, X.; and Li, S. 2023. Deep Neural Network Watermarking against Model Extraction Attack. In ACM MM

  37. [45]

    Z.; and Chen, Y

    Tang, M.; Dai, A.; DiValentin, L.; Ding, A.; Hass, A.; Gong, N. Z.; and Chen, Y. 2024. MODELGUARD: Information-Theoretic Defense Against Model Extraction Attacks. In USENIX Security Symposium

  38. [46]

    Tang, R.; Du, M.; Liu, N.; Yang, F.; and Hu, X. 2020. An embarrassingly simple approach for trojan attack in deep neural networks. In ACM SIGKDD

  39. [47]

    Wah, C.; Branson, S.; Welinder, P.; Perona, P.; and Belongie, S. 2011. The caltech-ucsd birds-200-2011 dataset

  40. [48]

    Wang, Y.; Li, J.; Liu, H.; Wang, Y.; Wu, Y.; Huang, F.; and Ji, R. 2022. Black-Box Dissector: Towards Erasing-based Hard-Label Model Stealing Attack. In ECCV

  41. [49]

    Wang, Y.; and Lin, X. 2022. Enhance model stealing attack via label refining. In ICSP. IEEE

  42. [50]

    Yu, H.; Yang, K.; Zhang, T.; Tsai, Y.-Y.; Ho, T.-Y.; and Jin, Y. 2020. CloudLeak: Large-Scale Deep Learning Models Stealing Through Adversarial Examples. In NDSS

  43. [51]

    Yuan, X.; Chen, K.; Huang, W.; Zhang, J.; Zhang, W.; and Yu, N. 2024. Data-Free Hard-Label Robustness Stealing Attack. In AAAI

  44. [52]

    Zhao, S.; Chen, K.; Hao, M.; Zhang, J.; Xu, G.; Li, H.; and Zhang, T. 2023. Extracting Cloud-based Model with Prior Knowledge. arXiv preprint arXiv:2306.04192

  45. [53]

    Zhao, Y.; Deng, X.; Liu, Y.; Pei, X.; Xia, J.; and Chen, W. 2024. Fully Exploiting Every Real Sample: SuperPixel Sample Gradient Model Stealing. In CVPR

  46. [54]

    Zhou, M.; Wu, J.; Liu, Y.; Liu, S.; and Zhu, C. 2020. DaST: Data-free Substitute Training for Adversarial Attacks. In CVPR

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.