End-to-End Cryptanalytic Extraction of Convolutional Neural Networks in the Hard-Label Setting

Chun LI , Zheng GONG , Di LI , Liping ZHUANG , Yufeng TANG , Yin LV , Xingfu YAN

Front. Comput. Sci. ››

PDF (3953KB)
Front. Comput. Sci. ›› DOI: 10.1007/s11704-026-61331-0
RESEARCH ARTICLE
End-to-End Cryptanalytic Extraction of Convolutional Neural Networks in the Hard-Label Setting
Author information +
History +
PDF (3953KB)

Abstract

Convolutional neural network (CNN) parameters are valuable intellectual property. Many application programming interfaces (APIs) return only top-1 labels to limit parameter recovery. Prior cryptanalytic extraction can recover functionally equivalent multilayer perceptrons (MLPs) with rectified linear unit (ReLU) activations. CNNs additionally require handling weight sharing, position-dependent embeddings of shared kernels, coupled spatial perturbations, and channel-sign ambiguity. This paper presents an end-to-end hard-label extraction attack for known-architecture ReLU CNN classifiers with average pooling. The main algorithmic contribution is conditional polynomial-time channel-sign recovery using singular-value gap ratio (SVGR)-guided retained-candidate discrete optimization. Given recovered unsigned signatures and valid, correctly attributed constraints, the polynomial-time guarantee applies only to sign recovery under polynomially bounded voting work, cumulative candidate checks across all layers and search rounds, and per-check work. The attack locates dual points on decision and activation boundaries, recovers shared channel signatures with singular value decomposition (SVD), resolves channel signs, and peels layers while absorbing ReLU scale factors into later linear layers. The 1D MNIST, 2D MNIST, and RGB CIFAR-10 experiments report 100% prediction fidelity. Further studies examine pooling placement and recursive numerical recovery through eleven convolutional layers. In the security evaluation, extracted watermarked CNNs retain behavior-level ownership evidence, and deterministic wrapper triggers introduce new behavior without erasing those signals.

Keywords

model extraction / hard-label oracle / convolutional neural networks / cryptanalytic attacks / AI security

Cite this article

Download citation ▾
Chun LI, Zheng GONG, Di LI, Liping ZHUANG, Yufeng TANG, Yin LV, Xingfu YAN. End-to-End Cryptanalytic Extraction of Convolutional Neural Networks in the Hard-Label Setting. Front. Comput. Sci. DOI:10.1007/s11704-026-61331-0

登录浏览全文

4963

注册一个新账户 忘记密码

References

Rights & permissions

Higher Education Press

PDF (3953KB)

0

Accesses

0

Citation

Detail

Sections
Recommended

/

〈 〉