Paper 2026/2850

Neural Cipher Bench: Cost and Portability of Corner-ReLU Cipher Implementation

Siwoo Eum, Hansung University
Minwoo Lee, Hansung University
Gyeongju Song, Hansung University
Seungwon Lee, Hansung University
Uijae Kim, Hansung University
Doyun Park, Hansung University
Minseo Kim, Hansung University
Wai Kong Lee, Universiti Tunku Abdul Rahman
Zhe Liu, Zhejiang University
Anupam Chattopadhyay, Nanyang Technological University
Hwajeong Seo, Hansung University
Abstract

Edge and Internet of Things (IoT) devices run neural networks on GPUs, neural processing units (NPUs), and microcontrollers (MCUs) through machine learning (ML) runtimes. Deep Neural Cryptography (DNC, EUROCRYPT 2026) expressed AES as a bit-exact ReLU network with fixed weights, using a corner-ReLU construction that keeps the cipher inside the affine+ReLU operators such runtimes execute. What this construction costs on edge and IoT hardware, and whether real ML backends execute it correctly, had not been measured. We implement SM4, ARIA, Camellia, and ChaCha20 (spanning substitution–permutation, Feistel, and add–rotate–XOR structures) alongside AES on a shared corner-ReLU core under one construction policy, derive a cost model for width, depth, and neuron count, and measure five execution paths on four platforms. All ciphers admit bit-exact networks whose width and depth diverge with cipher structure, and the model's width predicted where GPU throughput drops. Every configuration passing the correctness check fell below its classical baseline (best 0.80×, SM4 on the GPU), by hundreds to thousands of times for the S-box ciphers on the MCU, and keeping the activations on chip through fusion raised throughput 1.4–3.5× and moved the peak past the predicted cache limit without closing the gap. Some AES configurations on the Apple Neural Engine returned valid-looking but incorrect outputs with no error or warning, a failure reproduced in a minimal graph without project code. Mathematical bit-exactness therefore does not guarantee backend-level bit-exactness, and exact workloads offloaded to edge NPUs require output validation.

Metadata
Available format(s)
PDF
Category
Implementation
Publication info
Preprint.
Keywords
benchmarkingBit-exact executionDeep neural cryptographyNeural network hardwareSymmetric cryptography
Contact author(s)
shuraatum @ gmail com
minunejip @ gmail com
thdrudwn98 @ gmail com
dkajdfhd1 @ gmail com
chrisvt424 @ gmail com
rhcp030418 @ gmail com
msqoo0212 @ gmail com
waikong lee @ gmail com
zhe liu @ nuaa edu cn
anupam @ ntu edu sg
hwajeong84 @ gmail com
History
2026-11-12: approved
2026-11-10: received
See all versions
Short URL
https://ia.cr/2026/2850
License
Creative Commons Attribution
CC BY

BibTeX

@misc{cryptoeprint:2026/2850,
      author = {Siwoo Eum and Minwoo Lee and Gyeongju Song and Seungwon Lee and Uijae Kim and Doyun Park and Minseo Kim and Wai Kong Lee and Zhe Liu and Anupam Chattopadhyay and Hwajeong Seo},
      title = {Neural Cipher Bench: Cost and Portability of Corner-{ReLU} Cipher Implementation},
      howpublished = {Cryptology {ePrint} Archive, Paper 2026/2850},
      year = {2026},
      url = {https://eprint.iacr.org/2026/2850}
}
Note: In order to protect the privacy of readers, eprint.iacr.org does not use cookies or embedded third party content.