多路径特征增强的深度学习矩形攻击防御方法

doi:10.3969/j.issn.1671-1122.2026.07.009

摘要/Abstract

摘要：

针对传统矩形攻击与深度学习结合时存在的输入特征单一、密码特征识别能力受限的问题，文章以SPECK32/64算法及其变体SPECKEY算法为研究对象，提出一种扩展矩形神经网络差分区分器模型。该模型通过构造包含三重约束的密文数据组，将单向差分传播路径扩展为多路径特征输入，从而增强神经网络对高轮数密码算法非随机统计特征的捕捉能力。实验结果表明，基于该模型构建的5~8轮SPECK32/64区分器及6~8轮SPECKEY区分器，在分类准确率方面优于传统矩形神经网络区分器。此外，利用训练获得的8轮扩展矩形神经网络差分区分器，对10轮SPECK32/64算法进行子密钥恢复攻击。攻击结果显示，虽然完整16比特子密钥的完全恢复率为12%，但在剔除两个非敏感比特子密钥的影响后，模型对剩余14比特子密钥的预测准确率达100%，证明了增加多路径特征输入的有效性，为深度学习在密码分析领域的应用提供更具有效性的分析模型。

关键词: 深度学习, 矩形攻击, 神经网络, 密钥恢复攻击

Abstract:

To address the issues of single input features and limited capability in identifying cryptographic characteristics when combining traditional rectangle attacks with deep learning, this paper proposed an extended rectangle neural differential distinguisher model, taking the SPECK32/64 algorithm and its variant SPECKEY as the research objects. By constructing ciphertext data groups with triple constraints, the model expanded the unidirectional differential propagation path into multi-path feature inputs, thereby enhancing the neural network’s ability to capture non-random statistical characteristics of high-round ciphers. Experimental results show that the distinguishers for 5 to 8-round SPECK32/64 and 6 to 8-round SPECKEY construct based on this model all outperform the traditional rectangle neural distinguisher in classification accuracy. Furthermore, using the trained 8-round extended rectangle neural differential distinguisher, a subkey recovery attack is performed on the 10-round SPECK32/64 algorithm. The attack results show that while the complete recovery rate for the full 16-bit subkey is 12%, the model achieves a 100% prediction accuracy on the remaining 14 bits after eliminating the influence of two insensitive bits. This study demonstrates the effectiveness of incorporating multi-path feature inputs, providing a more effective analytical model for the application of deep learning in the field of cryptanalysis.

Key words: deep learning, rectangle attack, neural network, key recovery attack

中图分类号:

TP309

孙浩然, 陈杰, 刘君. 多路径特征增强的深度学习矩形攻击防御方法[J]. 信息网络安全, 2026, 26(7): 1115-1127.

Sun Haoran, Chen Jie, Liu Jun. Multi-path feature enhanced deep learning method for rectangule attack and defense[J]. Netinfo Security, 2026, 26(7): 1115-1127.

图/表 18

表1

图1

图2

图3

图4

图5

表2

图6

图7

图8

图9

表3

表4

表5

图10

图11

图12

图13

参考文献 28

[1]	Rivest R L. Cryptography and machine learning[C]// Cryptology-ASIACRYPT ’91. Heidelberg: Springer, 1993: 427-439.
[2]	Bafghi A G, Safabakhsh R, Sadeghiyan B. Finding the differential characteristics of block ciphers with neural networks[J]. Information Sciences, 2008, 178(15): 3118-3132. doi: 10.1016/j.ins.2008.02.016 URL
[3]	Soos M, Nohl K, Castelluccia C. Extending SAT solvers to cryptographic problems[C]// Theory and Applications of Satisfiability Testing-SAT 2009. Heidelberg: Springer, 2009: 244-257.
[4]	Mouha N, Preneel B. Towards finding optimal differential characteristics for ARX:application to salsa20[EB/OL]. (2013-06-02)[2026-04-25]. https://eprint.iacr.org/2013/328.
[5]	Bost R, Popa R A, Tu S, et al. Machine learning classification over encrypted data[EB/OL]. (2015-02-08)[2026-04-25]. https://www.ndss-symposium.org/ndss2015/ndss-2015-programme/machine-learning-classification-over-encrypted-data/.
[6]	De M F L, Xexeo J A M. Identifying encryption algorithms in ECB and CBC modes using computational intelligence[J]. Journal of Universal Computer Science, 2018, 24(1): 25-42.
[7]	Hu Xinyi, Zhao Yaqun. Research on plaintext restoration of AES based on neural network[EB/OL]. (2018-11-18)[2026-04-25]. https://onlinelibrary.wiley.com/doi/10.1155/2018/6868506.
[8]	Gomez A N, Huang Sicong, Zhang I, et al. Unsupervised cipher cracking using discrete GANs[EB/OL]. (2018-01-15)[2026-04-25]. https://arxiv.org/abs/1801.04883.
[9]	Gohr A. Improving attacks on round-reduced SPECK32/64 using deep learning[C]// Cryptology-CRYPTO 2019. Heidelberg: Springer, 2019: 150-179.
[10]	Jain A, Kohli V, Mishra G. Deep learning based differential distinguisher for lightweight block ciphers[EB/OL]. (2021-12-09)[2026-04-25]. https://arxiv.org/abs/2112.05061.
[11]	Zhang Liu, Wang Zilong. Improving differential-neural distinguisher model for DES, chaskey, and present[EB/OL]. (2022-04-12)[2026-04-25]. https://eprint.iacr.org/2022/457.
[12]	Yadav T, Kumar M. ML based improved differential distinguisher with high accuracy: application to GIFT-128 and ASCON[C]// Security, Privacy, and Applied Cryptography Engineering. Heidelberg: Springer, 2025: 287-316.
[13]	Baksi A, Breier J, Chen Yi, et al. Machine learning assisted differential distinguishers for lightweight ciphers[C]// 2021 Design, Automation & Test in Europe Conference & Exhibition (DATE). New York: IEEE, 2021: 176-181.
[14]	Bao Zhenzhen, Lu Jinyu, Yao Yiran, et al. More insight on deep learning-aided cryptanalysis[C]// Cryptology-ASIACRYPT 2023. Heidelberg: Springer, 2023: 436-467.
[15]	Yuan Xue, Wang Qichun. Improving differential-neural distinguisher for simeck family[EB/OL]. (2024-12-12)[2026-04-25]. https://eprint.iacr.org/2024/2002.
[16]	Yuan Xue, Wang Qichun. A multi-differential approach to enhance related-key neural distinguishers[EB/OL]. (2025-04-17)[2026-04-25]. https://eprint.iacr.org/2025/697.
[17]	So J. Deep learning-based cryptanalysis of lightweight block ciphers[EB/OL]. (2020-07-13)[2026-04-25]. https://onlinelibrary.wiley.com/doi/full/10.1155/2020/3701067.
[18]	Zhang Liu, Lu Jinyu, Wang Zilong, et al. Improved differential-neural cryptanalysis for round-reduced SIMECK32/64[EB/OL]. (2023-12-02)[2026-04-25]. https://link.springer.com/article/10.1007/s11704-023-3261-z.
[19]	Benamira A, Gerault D, Peyrin T, et al. A deeper look at machine learning-based cryptanalysis[C]// Cryptology-EUROCRYPT 2021. Heidelberg: Springer, 2021: 805-835.
[20]	Idris M F, Teh J S, Yan J L S, et al. A deep learning approach for active S-box prediction of lightweight generalized feistel block ciphers[J]. IEEE Access, 2021(9): 104205-104216.
[21]	Zhang Liu, Yao Yiran, Shi Danping, et al. Neural-inspired advances in integral cryptanalysis[C]// Cryptology-EUROCRYPT 2026. Heidelberg: Springer, 2026: 451-481.
[22]	Beaulieu R, Shors D, Smith J, et al. The SIMON and SPECK lightweight block ciphers[C]// The 52nd Annual Design Automation Conference. New York: ACM, 2015: 1-6.
[23]	Biryukov A, Velichkov V, Le C Y. Automatic search for the best trails in ARX: application to block cipher SPECK[C]// Fast Software Encryption. Heidelberg: Springer, 2016: 289-310.
[24]	Biham E, Dunkelman O, Keller N. The rectangle attack-rectangling the serpent[C]// Cryptology-EUROCRYPT 2001. Heidelberg: Springer, 2001: 340-357.
[25]	Chen Yi, Shen Yantian, Yu Hongbo, et al. A new neural distinguisher considering features derived from multiple ciphertext pairs[J]. Computer Journal, 2023, 66(6): 1419-1433.
[26]	孙浩然, 栗琳轲, 陈杰, 等. 针对轻量级分组密码的矩形攻击深度学习模型[J]. 西安电子科技大学学报, 2025, 52(6): 169-187.
[27]	Gohr A, Leander G, Neumann P P. An assessment of differential-neural distinguishers[EB/OL]. (2022-11-03)[2026-04-25]. https://eprint.iacr.org/2022/1521.
[28]	Zhang Liu, Wang Zilong, Wang Baocang. Improving differential-neural cryptanalysis[EB/OL]. (2024-10-07)[2026-04-25]. https://cic.iacr.org/p/1/3/13.

参数	值
优化器	Adam
损失函数	MSE
正则化参数	$5\times {{10}^{-7}}$
初始学习率	$2.2\times {{10}^{-4}}$
最大学习率	$3.5\times {{10}^{-3}}$
卷积核	$3\times 3$
批量大小/个	5000
训练集样本数/个	${{10}^{7}}$
验证集样本数/个	${{10}^{6}}$
测试集样本数/个	${{10}^{6}}$
训练周期数/轮	30
残差块数/个	10
循环学习率变化周期/轮	10

加密轮数/轮	文献来源	密文对数量/个	训练准确率
5	文献[25]	2 \| 4	97.39% \| 99.14%
	文献[26]	2 \| 4	99.83% \| 99.98%
	文献[27]	2 \| 4	97.80% \| 99.70%
	本文	3	99.97%
6	文献[25]	2 \| 4	86.67% \| 95.38%
	文献[26]	2 \| 4	96.53% \| 98.81%
	文献[27]	2 \| 4	87.70% \| 95.00%
	文献[28]	2 \| 4	87.73% \| 94.97%
	本文	3	98.93%
7	文献[25]	2 \| 4	63.96% \| 68.47%
	文献[26]	2 \| 4	75.37% \| 81.91%
	文献[27]	2 \| 4	66.30% \| 72.50%
	文献[28]	2 \| 4	66.49% \| 72.83%
	本文	3	89.06%
8	文献[26]	2 \| 4	53.19% \| 54.13%
	文献[27]	2 \| 4	52.10% \| 53.00%
	文献[28]	4	54.28%
	本文	3	65.98%

加密轮数/轮	训练准确率
6	99.04%
7	94.27%
8	77.14%

密码算法	加密轮数/轮	准确率	真阳性率	真阴性率
SPECK32/64	5	99.97%	99.96%	99.98%
	6	98.93%	98.61%	99.25%
	7	89.10%	85.42%	92.78%
	8	65.90%	62.99%	68.81%
SPECKEY	6	99.89%	98.11%	99.99%
	7	94.25%	92.12%	96.38%
	8	77.03%	68.23%	85.84%