CNN: Reverse Engineering based Extraction of Convolutional Neural Networks
IPSI transactions on Internet Research · 01.10.2026 · 22 (4) · 21-31
DOI: 10.58245/ipsi.tir.2604.03 ↗Kurz gefasst
Fachartikel in der Zeitschrift IPSI transactions on Internet Research, online erschienen am 01.10.2026. Verfasst von Lusky Yehonatan und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Neural network extraction attacks pose a significant threat to AI systems' security and intellectual property, enabling adversaries to replicate proprietary models, compromise confidentiality, or exploit functionality. In this paper, we present CopyCNN, a novel attack that extracts the exact weights and architecture of black-box, ReLU-based convolutional neural networks (CNNs) given the ability to issue arbitrary queries and observe the raw output logits (the network’s real-valued outputs prior to the softmax). To the best of our knowledge, CopyCNN is the first reverse engineering-based method capable of recovering precise weights from black-box CNNs. CopyCNN reverse engineers the core building blocks of modern CNNs—fully connected, convolutional, pooling, batch normalization, and residual layers. We establish the theoretical basis for CopyCNN and demonstrate its effectiveness through the successful extraction of several architectures. Our results highlight the practicality of CopyCNN for convolutional neural network extraction and its implications for IP protection, AI security, and computer vision applications.
Zitieren
APA
Yehonatan, L., Mendelson, A., & Nemcovsky, Y. (2026). CNN: Reverse Engineering based Extraction of Convolutional Neural Networks. IPSI transactions on Internet Research, 22(4), 21-31. https://doi.org/10.58245/ipsi.tir.2604.03
Vancouver
Yehonatan L, Mendelson A, Nemcovsky Y. CNN: Reverse Engineering based Extraction of Convolutional Neural Networks. IPSI transactions on Internet Research. 2026;22(4):21-31. doi:10.58245/ipsi.tir.2604.03
BibTeX
@article{Yehonatan2026,
author = {Yehonatan, Lusky and Mendelson, Avi and Nemcovsky, Yaniv},
title = {CNN: Reverse Engineering based Extraction of Convolutional Neural Networks},
journal = {IPSI transactions on Internet Research},
year = {2026},
volume = {22},
number = {4},
pages = {21--31},
doi = {10.58245/ipsi.tir.2604.03}
}
EUVIMED