acceptodds
Under review as a conference paper at ICLR 2027

Architecture-Agnostic Hard-Label Cryptanalytic Extraction of CNNs

Abstract

Cryptanalytic model extraction attacks aim to recover the exact weights of a neural network given only black-box access and partial architectural knowledge, posing significant risks to model confidentiality and intellectual property. Prior cryptanalytic attacks against convolutional neural networks (CNNs) have demonstrated exact recovery in the hard-label setting, but require detailed architectural knowledge, including the number, size, and types of each layer, channel counts, kernel sizes, strides, and padding. These requirements limit their applicability in realistic black-box settings. In this work, we show that such detailed architectural knowledge is not necessary by considering a significantly weaker threat model in which the attacker does not know any additional architecture knowledge besides what is necessary to query the model. Under this threat model, we develop procedures to recover convolutional hyperparameters directly from extracted weight blocks, infer channel counts, account for sequential linear layers such as average pooling and batch normalization, and extract convolutional layers that are followed by max-pool operations. Our results demonstrate that exact cryptanalytic extraction of CNNs remains feasible in this architecture agnostic setting, narrowing the gap between prior cryptanalytic extraction attacks and practical black-box extraction settings.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Model extraction, cryptanalytic extraction, hard-label, black-box, convolutional neural networks, max pooling, batch normalization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RK3AOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper extends a hard-label cryptanalytic extraction method to CNNs without requiring prior per-layer widths or convolution/pooling hyperparameters (the input shape is still assumed). Building on dual-point collection and clustering, it infers kernel shape, stride, padding, channel count, and max-pooling parameters from recovered weight vectors. Across five small CNNs, the paper reports exact architecture recovery for 6 of 8 attacked layers. The experiments use model gradients to obtain decision-boundary normals and assume the preceding layers are known.

Critical strengths
  1. The method removes several architecture assumptions made by prior exact CNN extraction attacks. Inferring convolution settings and channel counts from recovered weights, while extending the attack to max pooling, is a useful contribution.
  2. The evaluation includes varied cases such as 1D convolution, non-square kernels, folded average pooling and batch normalization, and max pooling. It reports incomplete recovery and runtime as well as successful cases, helping readers judge where the method works.
Critical weaknesses
  1. The experiments do not establish the full hard-label black-box result. The paper computes decision-boundary normals from model gradients and evaluates layers with the true preceding-layer prefix. These choices are disclosed, and end-to-end extraction is left open, but the results demonstrate layer recovery in an idealized setting rather than a complete label-only attack. This limits the practical claim.
  2. The study covers five small models and eight attacked layers. Six of the eight layers are recovered with their architecture exactly. The second non-square layer is assigned padding (1, 1), whereas the stated target architecture uses padding (2, 1), so its padding height is inferred as 1 rather than 2. The second LeNet-5 layer recovers only 1 of 16 unique kernels. These results support feasibility on the evaluated examples, but do not establish that the approach scales to larger or modern CNNs or that the full end-to-end attack works under the stated black-box threat model.
Questions for authors

Besides addressing the above mentioned weaknesses, address the following:

  1. Which steps in the reported experiments use gradients or other internal model access, and which use only hard-label queries? Do you have results for any complete layer using hard-label access alone?
  2. The supplement mentions each model was evaluated over 10 runs, while the tables show single values. Are these values aggregates or individual runs? Please clarify the aggregation and variability in coverage and dual budgets.
4 comments
AuthorsOpenReview ↗

Reviewer RK3A - Part 1/3

Thank you for your thoughtful feedback and constructive suggestions. We appreciate your insights and address your questions and concerns below.

R3-W1: Not full black-box end-to-end extraction

Response: Inspired by your insight, we have refined our attack to allow for full black-box end-to-end extraction in the hard-label setting. Our experiments no longer require perfect knowledge of previous layers, model gradients, or ground-truth neuron signs. We updated our evaluation on six representative models: 1D (2 Conv layers), Non-Square (2 Conv layers), Folded (L0: Conv, L1: AvgPool->Conv->BatchNorm), 2max-0fc (2 Conv->MaxPool layers), 2max-1fc (2 Conv->MaxPool followed by 1 FC), and LeNet-5 (2 Conv->MaxPool followed by 2 FC). All models are trained on MNIST with an input size of (1x28x28), except for 1D (3x80, normal distribution). We successfully extracted each model end-to-end to high fidelity without any white-box knowledge.

Table 4: End-to-end extraction results.

ModelFidelityDualsQueriesAttack time
1D100%10K8.13×1078.13{\times}10^744s
Non-Square100%10K1.73×1081.73{\times}10^8106s
Folded100%10K2.63×1082.63{\times}10^810m
2max-0fc100%20K4.77×1084.77{\times}10^86.7m
2max-1fc100%550K1.80×10101.80{\times}10^{10}4.4h
LeNet-599.84%4.5M1.6×10111.6{\times}10^{11}61h

All models were recovered with high fidelity (>99%), with 5/6 models recovered with 100% fidelity. The missing 0.16% fidelity in LeNet-5 is caused by missing near-dead neurons in its fully connected layers.

Revision: We added Section 5.3 and Table 4 with end-to-end extraction results for each model. We removed all references to knowledge of previous layer parameters and model gradients, and we removed end-to-end from future work.

R3-W2: Larger models

Response: To evaluate our approach on more complex CNNs and beyond image classification, we added experiments with variants of the HiDDeN decoder (HiDDeN-4, HiDDeN-8), a CNN designed to recover hidden messages from images, trained on CIFAR-10. These models have nine layers (eight hidden Conv->BN->ReLU layers followed by the model head), almost twice the depth of LeNet-5, and larger inputs (3×32×32 vs. 1×28×28).

Table 5: CNN Extraction Beyond LeNet-5.

ModelInput SizeDepthTotal NeuronsExtraction Result
LeNet-51×28×285 layers6508Full Extraction
HiDDeN-43×32×329 layers32768Full extraction
HiDDeN-83×32×329 layers61440First layer extracted. Full extraction can be done given additional time budget.

We successfully extracted all nine layers of HiDDeN-4 using approximately 500K dual points and the first layer of HiDDeN-8 using approximately 57K dual points within the past few days. For the three wider variants (HiDDeN-16, HiDDeN-32, and HiDDeN), our method also identified dual points with usable ASVs, with retention rates ranging from 41% to 63%. Given the computational and time cost of dual-point collection, we prioritized completing the extraction of HiDDeN-4 and extending layer recovery to HiDDeN-8, while evaluating dual-point discovery across all five variants. These results extend our demonstrated extraction capability beyond LeNet-scale image classifiers and provide initial evidence of its applicability to CNNs in other domains. The experiments across increasingly wider HiDDeN variants further establish a concrete direction for scaling our approach to more complex architectures. Although scaling to significantly more complex architectures remains an open challenge to the community, we believe that our approach lays a solid foundation. We have incorporated new experimental results and carefully calibrated our claims regarding model size and complexity to align with the evidence.

Revision: We updated Appendix B with the details of each HiDDeN variant. We added Section 5.4 and Table 5 with extraction of larger model sizes.

AuthorsOpenReview ↗

Reviewer RK3A - Part 2/3

R3-Q1: White-box access and per-layer results

Response: As explained in the response to R1-W1, we have refined our attack to allow for full black-box end-to-end extraction, and thus no steps use gradients or any white-box access. We successfully extracted all layers with our refined attack, including convolutional, fully connected, and the model heads. Table 3, shown below, reports per-layer results for complete end-to-end black-box extraction in the hard-label setting.

Table 3: Per-layer extraction metrics. Coverage is the fraction of a search space S on which no unrecovered effective weight is used (Liu et al., 2026). The reported number of duals is the number of additional duals required to extract the target layer.

ModelLnnCoverageMax Weight ErrorMax Bias ErrorDuals (#)Dual (s)Hyperparameter (s)Signature (s)Sign (s)
1D06/6100%8.9×10−168.9{\times}10^{-16}2.1×10−62.1{\times}10^{-6}10K340.80.340.05
11/1100%5.6×10−85.6{\times}10^{-8}2.7×10−52.7{\times}10^{-5}001.50.013.4
210/10100%2.0×10−82.0{\times}10^{-8}7.6×10−47.6{\times}10^{-4}-1.10.480.222.0
Non-Square04/4100%1.6×10−151.6{\times}10^{-15}3.4×10−73.4{\times}10^{-7}10K671.30.110.14
11/1100%6.7×10−166.7{\times}10^{-16}7.5×10−77.5{\times}10^{-7}005.60.114.2
210/10100%8.4×10−148.4{\times}10^{-14}1.8×10−51.8{\times}10^{-5}-7.50.153.916
Folded03/3100%0.05.5×10−85.5{\times}10^{-8}10K1301.61.40.06
13/3100%3.9×10−123.9{\times}10^{-12}3.7×10−63.7{\times}10^{-6}00550.19132
210/10100%5.7×10−95.7{\times}10^{-9}7.9×10−37.9{\times}10^{-3}-8.10.5321251
2max-0fc03/3100%7.8×10−167.8{\times}10^{-16}1.8×10−71.8{\times}10^{-7}10K1149.40.260.09
13/3100%3.1×10−153.1{\times}10^{-15}3.3×10−63.3{\times}10^{-6}10K2483.10.544.6
210/10100%3.9×10−153.9{\times}10^{-15}5.6×10−55.6{\times}10^{-5}-8.40.340.9114
2max-1fc03/3100%2.2×10−162.2{\times}10^{-16}4.8×10−74.8{\times}10^{-7}10K136110.300.08
18/8100%6.7×10−136.7{\times}10^{-13}4.7×10−44.7{\times}10^{-4}10K303120.932.5
255/6499.96%3.2×10−83.2{\times}10^{-8}4.3×10−34.3{\times}10^{-3}530K1499690167.7
310/1099.96%2.0×10−22.0{\times}10^{-2}1.61.6-7.71.10.05125
LeNet-506/6100%1.1×10−141.1{\times}10^{-14}1.6×10−51.6{\times}10^{-5}20K21391060.850.23
116/16100%6.7×10−156.7{\times}10^{-15}6.5×10−56.5{\times}10^{-5}480K82382126517294
296/12099.84%1.2×10−41.2{\times}10^{-4}1.6×10−31.6{\times}10^{-3}1.2M5448948119717
365/8485.61%1.0×10−11.0{\times}10^{-1}1.5×10−21.5{\times}10^{-2}2.8M7717848111529
410/1085.79%1.0×10−11.0{\times}10^{-1}4.4×10−14.4{\times}10^{-1}-8.30.557248

To help understand the numbers in Table 3, consider LeNet-5, where 96/120 neurons in L2 and 65/84 neurons in L3 were successfully extracted, achieving coverage (the fraction of a search space on which no unrecovered effective weight is used) of 99.84% and 85.61%, respectively. Most of the remaining neurons are dead or near-dead. Dead neurons do not contribute to the model's input-output behavior and thus have no impact on extraction fidelity. Near-dead neurons are activated only rarely and have limited impact on fidelity. Since our extraction process stops once the desired fidelity (99%) is achieved, some near-dead neurons remain unextracted, but they could be recovered with additional extraction time. Consequently, despite incomplete neuron recovery, the extracted LeNet-5 achieves 99.84% fidelity.

Max error generally grows with layer depth most visibly in 2max-1fc and LeNet-5, which contain FC layers after the initial convolutional layers. In 2max-1fc, 9 dead or near-dead neurons are missing in L2, lowering coverage from 100% to 99.96%, and increasing the error in the recovered head. For LeNet-5, 43 dead or near-dead neurons are missing in L2 and L3 combined, with high max error (0.1) and low coverage (<86%) across the last few model layers. The overall effect error accumulation recovered model is small, with both models extracted to high fidelity (>99%). Metrics for full end-to-end model extraction are reported in Table 4. Since coverage is measured as the fraction of the search space SS for which no unrecovered weight is used, low coverage is not a direct indicator of low fidelity.

Revision: We removed all references to knowledge of previous layer parameters and model gradients, and we removed end-to-end from future work. We updated Section 5.2 and Table 3 with per-layer results for each recovered model.

AuthorsOpenReview ↗

Reviewer RK3A - Part 3/3

R3-Q2: Multi-run aggregates

Response: We thank the reviewer for identifying this inconsistency. The statement that each model was evaluated over 10 runs was inadvertently retained from an earlier version of the manuscript and has now been corrected. The results reported in the main tables correspond to individual extraction runs, rather than averages across multiple runs. Therefore, the reported coverage and dual-point budgets are measurements from those individual runs, not aggregated statistics. To further examine variability, we have expanded our evaluation to include additional training configurations (Table 7 varies the seeds, Table 8 compares trained versus untrained models) of Folded and 2max-0fc, with the corresponding results reported in the Appendix. These additional experiments provide further insight into how extraction coverage and dual-point requirements vary across training configurations.

Table 7: Varied training seeds.

ModelSeedDualsDual Time (s)Queries
Folded4250K95311.2×1091.2{\times}10^9
6410K1052.1×1082.1{\times}10^8
2310K1182.2×1082.2{\times}10^8
1710K1202.3×1082.3{\times}10^8
8310K1362.4×1082.4{\times}10^8
2max-0fc4260K59281.7×1091.7{\times}10^9
6420K8595.0×1085.0{\times}10^8
2320K3874.8×1084.8{\times}10^8
1730K23717.7×1087.7{\times}10^8
83450K575341.4×10101.4{\times}10^{10}

The training seed occasionally affects the number of dual points and the time required to recover each model. For 2max-0fc, seed 83 required 450K duals to fully recover the second Conv+Pool layer. While extracting this specific layer, one kernel contained a low rate of dual RPCPs, which are required for a bias solution. After 450K dual points, this kernel contained two dual RPCPs, which is the minimum for a valid cluster. This indicates that neuron placement of trained models affects attack efficiency, especially in Conv+Pool layers, but not necessarily attack success.

Table 8: Trained versus untrained models.

ModelTrained?FidelityDualsDual Time (s)Queries
FoldedNo99.66%50K95311.2×1091.2{\times}10^9
Yes100%10K1052.1×1082.1{\times}10^8
2max-0fcNo99.92%60K59281.7×1091.7{\times}10^9
Yes100%20K8595.0×1085.0{\times}10^8

Table 8 reports a similar trend. The randomized neuron placement in the untrained models required more dual points to recover than the trained models. Additionally, the untrained models were extracted with lower fidelity, solely due to errors in the model head. Since these errors are within our fidelity threshold (>99%), we consider each model successfully extracted.

Revision: We added Appendix F and Table 7 with experiments showing varied training seeds. We added Appendix G and Table 8 with experiments showing trained versus untrained model extraction.

Dear Reviewer,

With the discussion period nearing its conclusion, we would like to express our sincere appreciation for your time and thoughtful feedback. We would greatly appreciate hearing whether our responses and additional experiments have adequately addressed your concerns. Should any questions or concerns remain, we would be happy to provide further clarification.

Thank you again for your valuable time and consideration.

The Authors

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

The paper extends existing cryptanalytic CNN extraction methods to recover architectural details that prior attacks assume to be known in the threat model. Starting from extracted neuron weight vectors, convolutional kernel sizes, strides, padding and output-channel counts are inferred. It also identifies max pooling and estimates its parameters, while treating consecutive affine operations such as convolution and batch normalisation as one equivalent layer. The evaluation covers five small CNN configurations and recovers the architecture exactly for six of eight tested layers. Recovery becomes substantially harder in deeper layers: only one of sixteen kernels is recovered in the second convolutional layer of the LeNet-style model. Overall, the results demonstrate that architectural information can be inferred from recovered weights for simple and small CNN models.

Critical strengths
  • Inferring kernel size, stride, padding and channel counts from recovered weights is an interesting extension of existing CNN extraction methods.
  • Reduces the architectural information the attacker needs beforehand.
  • Tests several configurations, including non-square kernels, max pooling and batch normalisation.
  • Reports recovery failures and computational costs, helping identify where the method struggles.
Critical weaknesses
  • Does not demonstrate end-to-end extraction: each layer is recovered with perfect knowledge of preceding layers.
  • Uses model gradients and ground-truth neuron signs in experiments, so the full hard-label attack is not tested.
  • Evaluates recovery only in the first two convolutional layers. For LeNet’s second layer, only 1 of 16 kernels is recovered.
  • Missing observations can produce incorrect padding, stride or channel counts. The paper does not show how these errors would be detected and corrected before extracting subsequent layers.
  • Supports a restricted family of sequential architectures with ReLU and pooling. Does not examine how training data, training seeds or the number of classes affect recovery success and cost.
  • The practical significance of cryptanalytical extraction of such small CNNs remains limited. This is the biggest reason for my scoring.
Questions for authors
  • How does recovery success vary across training seeds, trained versus untrained models, and real versus random-label training? Having some statistics on how practically usable this attack is on the same model architecture types but with above differences would be very interesting to gain some more intuition on neuron placement making the formulas solvable.
  • Does binary versus multiclass classification and the effects on neuron placement in the model affect the availability of informative dual points and recovery cost?
  • If one layer’s padding, stride or channel count is inferred incorrectly, can the method detect and correct the error?
  • How exactly are pooling-switch clusters identified when the two shifted kernels overlap, and how often are they misclassified?
  • Could you clarify how the shifted-kernel criterion is implemented when the kernel copies overlap, and report how accurately it distinguishes pooling-switch clusters from single-neuron clusters?
  • The method increases the number of dual points until the inferred architecture stops changing. How do you check whether all kernels and the boundary positions needed to infer padding have been observed, rather than repeatedly finding the same subset?
5 comments
AuthorsOpenReview ↗

Reviewer tbmN - Part 1/4

Thank you for your thoughtful feedback and constructive suggestions. We appreciate your insights and address your questions and concerns below.

R2-W1, R2-W2: Not full black-box end-to-end extraction

Response: Inspired by your insight, we have refined our attack to allow for full black-box end-to-end extraction in the hard-label setting. Our experiments no longer require perfect knowledge of previous layers, model gradients, or ground-truth neuron signs. We updated our evaluation on six representative models: 1D (2 Conv layers), Non-Square (2 Conv layers), Folded (L0: Conv, L1: AvgPool->Conv->BatchNorm), 2max-0fc (2 Conv->MaxPool layers), 2max-1fc (2 Conv->MaxPool followed by 1 FC), and LeNet-5 (2 Conv->MaxPool followed by 2 FC). All models are trained on MNIST with an input size of (1x28x28), except for 1D (3x80, normal distribution). We successfully extracted each model end-to-end to high fidelity without any white-box knowledge.

Table 4: End-to-end extraction results.

ModelFidelityDualsQueriesAttack time
1D100%10K8.13×1078.13{\times}10^744s
Non-Square100%10K1.73×1081.73{\times}10^8106s
Folded100%10K2.63×1082.63{\times}10^810m
2max-0fc100%20K4.77×1084.77{\times}10^86.7m
2max-1fc100%550K1.80×10101.80{\times}10^{10}4.4h
LeNet-599.84%4.5M1.6×10111.6{\times}10^{11}61h

All models were recovered with high fidelity (>99%), with 5/6 models recovered with 100% fidelity. The missing 0.16% fidelity in LeNet-5 is caused by missing near-dead neurons in its fully connected layers.

Revision: We added Section 5.3 and Table 4 with end-to-end extraction results for each model. We removed all references to knowledge of previous layer parameters and model gradients, and we removed end-to-end from future work.

R2-W3: Only first two convolutional layers, missing kernels in LeNet-5

Response: As seen in the last response, we have refined our attack to be end-to-end. Therefore, we now additionally recover fully connected layers, signs, and the model head. As part of this refinement, we optimized attack steps for efficiency, reducing the time cost per layer (the reported time cost in Table 3 reflect the efficiency improvements). For LeNet-5, this resulted in a full recovery of all 16 kernels in the second hidden layer. 

Table 3: Per-layer extraction metrics. Coverage is the fraction of a search space S on which no unrecovered effective weight is used (Liu et al., 2026). The reported number of duals is the number of additional duals required to extract the target layer.

ModelLnnCoverageMax Weight ErrorMax Bias ErrorDuals (#)Dual (s)Hyperparameter (s)Signature (s)Sign (s)
1D06/6100%8.9×10−168.9{\times}10^{-16}2.1×10−62.1{\times}10^{-6}10K340.80.340.05
11/1100%5.6×10−85.6{\times}10^{-8}2.7×10−52.7{\times}10^{-5}001.50.013.4
210/10100%2.0×10−82.0{\times}10^{-8}7.6×10−47.6{\times}10^{-4}-1.10.480.222.0
Non-Square04/4100%1.6×10−151.6{\times}10^{-15}3.4×10−73.4{\times}10^{-7}10K671.30.110.14
11/1100%6.7×10−166.7{\times}10^{-16}7.5×10−77.5{\times}10^{-7}005.60.114.2
210/10100%8.4×10−148.4{\times}10^{-14}1.8×10−51.8{\times}10^{-5}-7.50.153.916
Folded03/3100%0.05.5×10−85.5{\times}10^{-8}10K1301.61.40.06
13/3100%3.9×10−123.9{\times}10^{-12}3.7×10−63.7{\times}10^{-6}00550.19132
210/10100%5.7×10−95.7{\times}10^{-9}7.9×10−37.9{\times}10^{-3}-8.10.5321251
2max-0fc03/3100%7.8×10−167.8{\times}10^{-16}1.8×10−71.8{\times}10^{-7}10K1149.40.260.09
13/3100%3.1×10−153.1{\times}10^{-15}3.3×10−63.3{\times}10^{-6}10K2483.10.544.6
210/10100%3.9×10−153.9{\times}10^{-15}5.6×10−55.6{\times}10^{-5}-8.40.340.9114
2max-1fc03/3100%2.2×10−162.2{\times}10^{-16}4.8×10−74.8{\times}10^{-7}10K136110.300.08
18/8100%6.7×10−136.7{\times}10^{-13}4.7×10−44.7{\times}10^{-4}10K303120.932.5
255/6499.96%3.2×10−83.2{\times}10^{-8}4.3×10−34.3{\times}10^{-3}530K1499690167.7
310/1099.96%2.0×10−22.0{\times}10^{-2}1.61.6-7.71.10.05125
LeNet-506/6100%1.1×10−141.1{\times}10^{-14}1.6×10−51.6{\times}10^{-5}20K21391060.850.23
116/16100%6.7×10−156.7{\times}10^{-15}6.5×10−56.5{\times}10^{-5}480K82382126517294
296/12099.84%1.2×10−41.2{\times}10^{-4}1.6×10−31.6{\times}10^{-3}1.2M5448948119717
365/8485.61%1.0×10−11.0{\times}10^{-1}1.5×10−21.5{\times}10^{-2}2.8M7717848111529
410/1085.79%1.0×10−11.0{\times}10^{-1}4.4×10−14.4{\times}10^{-1}-8.30.557248
AuthorsOpenReview ↗

Reviewer tbmN - Part 2/4

R2-W3 (Continued)

To help understand the numbers in Table 3, consider LeNet-5, where 96/120 neurons in L2 and 65/84 neurons in L3 were successfully extracted, achieving coverage (the fraction of a search space on which no unrecovered effective weight is used) of 99.84% and 85.61%, respectively. Most of the remaining neurons are dead or near-dead. Dead neurons do not contribute to the model's input-output behavior and thus have no impact on extraction fidelity. Near-dead neurons are activated only rarely and have limited impact on fidelity. Since our extraction process stops once the desired fidelity (99%) is achieved, some near-dead neurons remain unextracted, but they could be recovered with additional extraction time. Consequently, despite incomplete neuron recovery, the extracted LeNet-5 achieves 99.84% fidelity.

Max error generally grows with layer depth most visibly in 2max-1fc and LeNet-5, which contain FC layers after the initial convolutional layers. In 2max-1fc, 9 dead or near-dead neurons are missing in L2, lowering coverage from 100% to 99.96%, and increasing the error in the recovered head. For LeNet-5, 43 dead or near-dead neurons are missing in L2 and L3 combined, with high max error (0.1) and low coverage (<86%) across the last few model layers. The overall effect of the error accumulation on recovered model is small, with both models extracted to high fidelity (>99%). Metrics for full end-to-end model extraction are reported in Table 4 above. Since coverage is measured as the fraction of the search space SS for which no unrecovered weight is used, low coverage is not a direct indicator of low fidelity.

Revision: We updated Section 5.2 and Table 3 with per-layer results for each recovered model.

R2-W4 (R2-Q3): Correcting padding, stride, and channel errors

Response: Our refined approach uses Approximate Signature Vectors (ASVs) to infer layer hyperparameters, followed by a verification and correction step for padding and stride. While incorrect channel counts are not directly detected or corrected, they can introduce large error that propagates to the model head, resulting in low overall extraction fidelity. Such low fidelity serves as an indicator of potential extraction errors, prompting the collection of additional dual points and re-extraction of the affected layers. Importantly, our experiments achieve consistently high extraction fidelity (>99%, Table 4), demonstrating the effectiveness of our recovery pipeline on the evaluated models.

Comparison Results: A comparison of our original submission results and our post-refinement results on three unchanged models. The number of duals only counts dual points used to extract the first two model layers.

Model# Duals (pre-refinement)Hyperparameters correct? (pre-refinement)# Duals (post-refinement)Hyperparameters correct? (post-refinement)
1D20KYes10KYes
Non-Square100KNo (Padding)10KYes
LeNet-5300KNo (1/16 channels)500KYes (16/16 channels)

As shown in the above comparison results, our refined approach uses significantly less dual points to recover convolutional layers, and recovers hyperparameters exactly. For 1D and non-square, the reduced number of dual points directly corresponds to the padding error correction during kernel-centric clustering. While the number of dual points grew in LeNet-5, all 16 of the second layer's channels are successfully extracted instead of just one.

Revision: We added clarification on correcting hyperparameters during kernel-centric clustering in Section 4.1.3.

AuthorsOpenReview ↗

Reviewer tbmN - Part 3/4

R2-W6: Larger model sizes

Response: To evaluate our approach on more complex CNNs and beyond image classification, we added experiments with variants of the HiDDeN decoder (HiDDeN-4, HiDDeN-8), a CNN designed to recover hidden messages from images, trained on CIFAR-10. These models have nine layers (eight hidden Conv->BN->ReLU layers followed by the model head), almost twice the depth of LeNet-5, and larger inputs (3×32×32 vs. 1×28×28).

Table 5: CNN Extraction Beyond LeNet-5.

ModelInput SizeDepthTotal NeuronsExtraction Result
LeNet-51×28×285 layers6508Full Extraction
HiDDeN-43×32×329 layers32768Full extraction
HiDDeN-83×32×329 layers61440First layer extracted. Full extraction can be done given additional time budget.

We successfully extracted all nine layers of HiDDeN-4 using approximately 500K dual points and the first layer of HiDDeN-8 using approximately 57K dual points within the past few days. For the three wider variants (HiDDeN-16, HiDDeN-32, and HiDDeN), our method also identified dual points with usable ASVs, with retention rates ranging from 41% to 63%. Given the computational and time cost of dual-point collection, we prioritized completing the extraction of HiDDeN-4 and extending layer recovery to HiDDeN-8, while evaluating dual-point discovery across all five variants. These results extend our demonstrated extraction capability beyond LeNet-scale image classifiers and provide initial evidence of its applicability to CNNs in other domains. The experiments across increasingly wider HiDDeN variants further establish a concrete direction for scaling our approach to more complex architectures. Although scaling to significantly more complex architectures remains an open challenge to the community, we believe that our approach lays a solid foundation. We have incorporated new experimental results and carefully calibrated our claims regarding model size and complexity to align with the evidence.

Revision: We updated Appendix B with the details of each HiDDeN variant. We added Section 5.4 and Table 5 with extraction of larger model sizes.

R2-Q1 (part of R2-W5): Varying training configurations

Response: Our attack successfully recovers models across various training configurations. We evaluate 5 training seeds on two models to efficiently assess full-model extraction. We observe that neuron placement primarily affects extraction efficiency, as certain neurons require substantially more dual points to recover, particularly when max-pooling is involved. Although neither untrained model was recovered with 100% fidelity, both reached our threshold of 99% fidelity. The remaining discrepancies stemmed from errors in extracting the output head rather than the intermediate layers.

Table 7: Varied training seeds.

ModelSeedDualsDual Time (s)Queries
Folded4250K95311.2×1091.2{\times}10^9
6410K1052.1×1082.1{\times}10^8
2310K1182.2×1082.2{\times}10^8
1710K1202.3×1082.3{\times}10^8
8310K1362.4×1082.4{\times}10^8
2max-0fc4260K59281.7×1091.7{\times}10^9
6420K8595.0×1085.0{\times}10^8
2320K3874.8×1084.8{\times}10^8
1730K23717.7×1087.7{\times}10^8
83450K575341.4×10101.4{\times}10^{10}

The training seed occasionally affects the number of dual points and the time required to recover each model. For 2max-0fc, seed 83 required 450K duals to fully recover the second Conv+Pool layer. While extracting this specific layer, one kernel contained a low rate of dual ReLU-Pooling Critical Points (RPCPs), which are required for a bias solution. After 450K dual points, this kernel contained two dual RPCPs, which is the minimum for a valid cluster. This indicates that neuron placement of trained models affects attack efficiency, especially in Conv+Pool layers, but not necessarily attack success.

Table 8: Untrained models.

ModelTrained?FidelityDualsDual Time (s)Queries
FoldedNo99.66%50K95311.2×1091.2{\times}10^9
Yes100%10K1052.1×1082.1{\times}10^8
2max-0fcNo99.92%60K59281.7×1091.7{\times}10^9
Yes100%20K8595.0×1085.0{\times}10^8

Table 8 reports a similar trend. The randomized neuron placement in the untrained models required more dual points to recover than the trained models. Additionally, the untrained models were extracted with lower fidelity, solely due to errors in the model head. Since these errors are within our fidelity threshold (>99%), we consider each model successfully extracted.

Revision: We added Appendix F and Table 7 with experiments showing varied training seeds. We added Appendix G and Table 8 with experiments showing trained versus untrained model extraction.

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

The paper shows that the architectural knowledge is not necessary by considering a significantly weaker threat model in which the attacker does not know any additional architecture knowledge besides what is necessary to query the model. Under this threat model, they develop procedures to recover convolutional hyperparameters directly from extracted weight blocks, infer channel counts, account for sequential linear layers such as average pooling and batch normalization, and extract convolutional layers that are followed by max-pool operations. The results show the effectiveness.

Critical strengths
  • The paper addresses an important topic in the research field.

  • The paper illustrates the relate work and background in detail.

Critical weaknesses
  • The paper says "In this paper, we focus on layer-wise functionally equivalent extraction. Each layer is recovered given perfect knowledge of previous layers." What if previous layers are extracted not given? If there are some errors in the extraction of previous layers, will these errors accumulate? In this case, what would be the whole performance?

  • The largest model in the paper is LeNet. LeNet is a pretty small model compared to models nowadays. What is the practical impact on today's research field? Can these conclusions apply to today's complex large models?

  • The paper says dual-point collection, clustering, and kernel recovery are all the same as previous paper, making this paper incremental.

Questions for authors

See comments.

4 comments
AuthorsOpenReview ↗

Reviewer EszW - Part 1/3

Thank you for your thoughtful feedback and constructive suggestions. We appreciate your insights and address your questions and concerns below.

R1-W1: Effects of error accumulation

Response: We have refined our attack to allow for full black-box end-to-end extraction in the hard-label setting. Attackers no longer assume perfect knowledge of previous layers, causing errors to propagate into deeper layers. We evaluate six representative models: 1D (2 Conv layers), Non-Square (2 Conv layers), Folded (L0: Conv, L1: AvgPool->Conv->BatchNorm), 2max-0fc (2 Conv->MaxPool layers), 2max-1fc (2 Conv->MaxPool followed by 1 FC), and LeNet-5 (2 Conv->MaxPool followed by 2 FC). All models are trained on MNIST with an input size of (1x28x28), except for 1D (3x80, normal distribution). We successfully recovered each model end-to-end using previously extracted layer's weights with accumulated errors to extract deeper layers. Per-layer extraction metrics using our refined approach can be found in Table 3 (in the revised paper).

Table 3: Per-layer extraction metrics. Coverage is the fraction of a search space SS on which no unrecovered effective weight is used (Liu et al., 2026). The reported number of duals is the number of additional duals required to extract the target layer.

ModelLnnCoverageMax Weight ErrorMax Bias ErrorDuals (#)Dual (s)Hyperparameter (s)Signature (s)Sign (s)
1D06/6100%8.9×10−168.9{\times}10^{-16}2.1×10−62.1{\times}10^{-6}10K340.80.340.05
11/1100%5.6×10−85.6{\times}10^{-8}2.7×10−52.7{\times}10^{-5}001.50.013.4
210/10100%2.0×10−82.0{\times}10^{-8}7.6×10−47.6{\times}10^{-4}-1.10.480.222.0
Non-Square04/4100%1.6×10−151.6{\times}10^{-15}3.4×10−73.4{\times}10^{-7}10K671.30.110.14
11/1100%6.7×10−166.7{\times}10^{-16}7.5×10−77.5{\times}10^{-7}005.60.114.2
210/10100%8.4×10−148.4{\times}10^{-14}1.8×10−51.8{\times}10^{-5}-7.50.153.916
Folded03/3100%0.05.5×10−85.5{\times}10^{-8}10K1301.61.40.06
13/3100%3.9×10−123.9{\times}10^{-12}3.7×10−63.7{\times}10^{-6}00550.19132
210/10100%5.7×10−95.7{\times}10^{-9}7.9×10−37.9{\times}10^{-3}-8.10.5321251
2max-0fc03/3100%7.8×10−167.8{\times}10^{-16}1.8×10−71.8{\times}10^{-7}10K1149.40.260.09
13/3100%3.1×10−153.1{\times}10^{-15}3.3×10−63.3{\times}10^{-6}10K2483.10.544.6
210/10100%3.9×10−153.9{\times}10^{-15}5.6×10−55.6{\times}10^{-5}-8.40.340.9114
2max-1fc03/3100%2.2×10−162.2{\times}10^{-16}4.8×10−74.8{\times}10^{-7}10K136110.300.08
18/8100%6.7×10−136.7{\times}10^{-13}4.7×10−44.7{\times}10^{-4}10K303120.932.5
255/6499.96%3.2×10−83.2{\times}10^{-8}4.3×10−34.3{\times}10^{-3}530K1499690167.7
310/1099.96%2.0×10−22.0{\times}10^{-2}1.61.6-7.71.10.05125
LeNet-506/6100%1.1×10−141.1{\times}10^{-14}1.6×10−51.6{\times}10^{-5}20K21391060.850.23
116/16100%6.7×10−156.7{\times}10^{-15}6.5×10−56.5{\times}10^{-5}480K82382126517294
296/12099.84%1.2×10−41.2{\times}10^{-4}1.6×10−31.6{\times}10^{-3}1.2M5448948119717
365/8485.61%1.0×10−11.0{\times}10^{-1}1.5×10−21.5{\times}10^{-2}2.8M7717848111529
410/1085.79%1.0×10−11.0{\times}10^{-1}4.4×10−14.4{\times}10^{-1}-8.30.557248

To help understand the numbers in Table 3, consider LeNet-5, where 96/120 neurons in L2 and 65/84 neurons in L3 were successfully extracted, achieving coverage (the fraction of a search space on which no unrecovered effective weight is used) of 99.84% and 85.61%, respectively. Most of the remaining neurons are dead or near-dead. Dead neurons do not contribute to the model's input-output behavior and thus have no impact on extraction fidelity. Near-dead neurons are activated only rarely and have limited impact on fidelity. Since our extraction process stops once the desired fidelity (99%) is achieved, some near-dead neurons remain unextracted, but they could be recovered with additional extraction time. Consequently, despite incomplete neuron recovery, the extracted LeNet-5 achieves 99.84% fidelity.

Max error generally grows with layer depth most visibly in 2max-1fc and LeNet-5, which contain FC layers after the initial convolutional layers. In 2max-1fc, 9 dead or near-dead neurons are missing in L2, lowering coverage from 100% to 99.96%, and increasing the error in the recovered head. For LeNet-5, 43 dead or near-dead neurons are missing in L2 and L3 combined, with high max error (0.1) and low coverage (<86%) across the last few model layers. The overall effect of the error accumulation on the recovered model is small, with both models extracted to high fidelity (>99%). Metrics for full end-to-end model extraction are reported in Table 4.

AuthorsOpenReview ↗

Reviewer EszW - Part 2/3

R1-W1 (Continued)

Table 4: End-to-end extraction results.

ModelFidelityDualsQueriesAttack time
1D100%10K8.13×1078.13{\times}10^744s
Non-Square100%10K1.73×1081.73{\times}10^8106s
Folded100%10K2.63×1082.63{\times}10^810m
2max-0fc100%20K4.77×1084.77{\times}10^86.7m
2max-1fc100%550K1.80×10101.80{\times}10^{10}4.4h
LeNet-599.84%4.5M1.6×10111.6{\times}10^{11}61h

All models were recovered with high fidelity (>99%), with 5/6 models recovered with 100% fidelity. Even though the larger two models both contained accumulated error in per-layer metrics, such as coverage below 100% in FC layers and high error in the model head, both models were recovered successfully. Since coverage is measured as the fraction of the search space SS for which no unrecovered weight is used, low coverage is not a direct indicator of low fidelity.

Revision: We updated Section 5.2 and Table 3 with per-layer results for each recovered model. We added Section 5.3 and Table 4 with end-to-end extraction results for each model. We removed all references to knowledge of previous layer parameters and model gradients, and we removed end-to-end from future work since it has been accomplished.

R1-W2: Larger model sizes

Response: To evaluate our approach on more complex CNNs and beyond image classification, we added experiments with variants of the HiDDeN decoder (HiDDeN-4, HiDDeN-8), a CNN designed to recover hidden messages from images, trained on CIFAR-10. These models have nine layers (eight hidden Conv->BN->ReLU layers followed by the model head), almost twice the depth of LeNet-5, and larger inputs (3×32×32 vs. 1×28×28).

Table 5: CNN Extraction Beyond LeNet-5.

ModelInput SizeDepthTotal NeuronsExtraction Result
LeNet-51×28×285 layers6508Full Extraction
HiDDeN-43×32×329 layers32768Full extraction
HiDDeN-83×32×329 layers61440First layer extracted. Full extraction can be done given additional time budget.

We successfully extracted all nine layers of HiDDeN-4 using approximately 500K dual points and the first layer of HiDDeN-8 using approximately 57K dual points within the past few days. For the three wider variants (HiDDeN-16, HiDDeN-32, and HiDDeN), our method also identified dual points with usable ASVs, with retention rates ranging from 41% to 63%. Given the computational and time cost of dual-point collection, we prioritized completing the extraction of HiDDeN-4 and extending layer recovery to HiDDeN-8, while evaluating dual-point discovery across all five variants. These results extend our demonstrated extraction capability beyond LeNet-scale image classifiers and provide initial evidence of its applicability to CNNs in other domains. The experiments across increasingly wider HiDDeN variants further establish a concrete direction for scaling our approach to more complex architectures. Although scaling to significantly more complex architectures remains an open challenge to the community, we believe that our approach lays a solid foundation. We have incorporated new experimental results and carefully calibrated our claims regarding model size and complexity to align with the evidence.

Revision: We updated Appendix B with the details of each HiDDeN variant. We added Section 5.4 and Table 5 with extraction of larger model sizes.

AuthorsOpenReview ↗

Reviewer EszW - Part 3/3

R1-W3: Reuse of prior techniques

Response: While we build upon existing techniques, our contribution is not simply improving these individual techniques, but developing a new extraction pipeline that enables capabilities much beyond those of prior work. 

First, our approach substantially relaxes the assumptions required by prior work. Rather than assuming knowledge of the victim model's architecture and convolutional hyperparameters, our approach builds the architecture while recovering the model's parameters, enabling model extraction under significantly more restrictive black-box conditions. 

Second, we introduce new techniques and adapt existing ones to support this expanded capability. Specifically, we identify layer hyperparameters from neuron-centric clusters, distinguish between layer types, and merge sequential linear layers into one affine transformation. Inspired by the reviewers' wonderful questions and suggestions, we further refined our attack by optimizing the dual-point search for the GPU, introducing targeted dual-point collection for deeper layers, identifying hyperparameters from Approximate Signature Vectors (ASVs) before finding kernel solutions, and applying kernel-centric clustering to verify stride and padding before jointly recovering kernels and biases. These components form an integrated pipeline that goes beyond simply reusing the techniques of prior work.  

Third, we enable end-to-end black-box extraction without assuming perfect recovery of preceding layers. Our approach tolerates error propagation across layers with minimal impact on overall extraction fidelity, making the attack more practical under realistic conditions. 

Fourth, we extend extraction to additional layer types, including Batch Normalization, average pooling, max pooling, and global average pooling. Batch Normalization and global average pooling were not supported by any previous extraction pipeline. 

Collectively, our contributions go beyond efficiency improvements: they extend the scope of cryptanalytic CNN extraction by reducing the required architectural knowledge, enabling error-tolerant end-to-end black-box extraction, introducing new recovery capabilities, and supporting additional layer types.

Revision: We updated Sections 4.6 and Algorithm 1 (now in Appendix A) with a full breakdown of a single layer extraction. We updated Section 3.2.1 and added Appendices C and D, explaining the optimizations to the dual point search and providing targeted search results. We updated Section 4.1 to describe ASV hyperparameter extraction, 3.2.2 to describe both clustering processes, 3.2.3 with the updated kernel recovery, and added 3.2.4 to describe sign recovery.

Dear Reviewer,

With the discussion period nearing its conclusion, we would like to express our sincere appreciation for your time and thoughtful feedback. We would greatly appreciate hearing whether our responses and additional experiments have adequately addressed your concerns. Should any questions or concerns remain, we would be happy to provide further clarification.

Thank you again for your valuable time and consideration.

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Appreciation and Summary

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their careful evaluation and thoughtful feedback, which has helped us refine our paper. We appreciate the reviewers' recognition of the significance of our architecture-agnostic approach. Hard-label cryptanalytic extraction is a difficult problem that demands substantial compute and engineering effort. We believe our work significantly relaxes attacker assumptions, allowing for full black-box end-to-end extraction in the hard-label, architecture-agnostic setting. We reduce time and query budgets through pipeline optimizations. We recover each evaluated model to high fidelity, and scale to larger and more complex architectures than prior work, including models with bit-decoder outputs, batch normalization, and global average pooling.

We have addressed all reviewers' questions and concerns in our individual responses. We have also uploaded a revised manuscript, with changes from the original submission highlighted in blue. Further details are provided in our reviewer-specific responses, and we would be happy to offer additional clarifications if needed.

Thank you again for your time, thoughtful comments, and valuable feedback.

Best,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.