一种融合多尺度卷积网络和自注意力结构的神经区分器OA
A Neural Distinguisher Integrating Multi-Scale Convolutional Network and Self-Attention Structure
针对传统卷积结构的神经区分器对密文数据分类精度较低的问题,提出一种融合卷积网络和自注意力结构的神经区分器.首先,通过设计一种多尺度卷积网络,以学习密文序列的局部特征;其次,在多尺度卷积网络后面引入自注意力机制,用来增强神经区分器捕捉密文序列中长距离依赖的能力;最后,在多种密文数据集上训练该神经区分器.实验结果表明,该神经区分器在8种密文数据集上的分类精度都超越了之前性能较优的神经区分器SE-ResNet,特别是在13轮的SIMON64上,该神经区分器的分类精度比SE-ResNet提高了3.92%,证明了该神经区分器具有更加强大的密文特征提取能力.
To address the low classification accuracy of traditional convolutional neural distinguishers on ciphertext data,neural distinguisher based on convolutional network and self-attention structure is proposed.Firstly,a multi-scale convolutional network is designed to learn local features of ciphertext sequences.Secondly,the self-attention mechanism is introduced after the multi-scale convolutional network to enhance the ability of the neural distinguisher to capture long-range dependencies in cipher-text sequences.Finally,the neural distinguisher is trained on multiple ciphertext datasets.Experimen-tal results show that the proposed neural distinguisher achieves higher classification accuracy than the previously superior neural distinguisher SE-ResNet on all eight ciphertext datasets.In particular,on 13-round SIMON64,the classification accuracy of the neural distinguisher is imperoved by 3.92%compared with that of SE-ResNet,demonstrating its more powerful ciphertext feature extraction capabil-ity.
任文欢;刘天源;王林元;陈健;闫镔
信息工程大学,河南 郑州 450001信息工程大学,河南 郑州 450001信息工程大学,河南 郑州 450001信息工程大学,河南 郑州 450001信息工程大学,河南 郑州 450001
信息技术与安全科学
多尺度卷积网络自注意力机制神经区分器密文序列差分密码分析
multi-scale convolutional networkself-attention mechanismneural distinguisherci-phertext sequencedifferential cryptanalysis
《信息工程大学学报》 2026 (4)
430-435,449,7
国家自然科学基金(62271504)
评论