基于深度学习的差分神经区分器求解方法  

Method for solving differential neural distinguishers based on deep learning

作　　者：蒋若怡 韦永壮[1] 王慧娇[1] JIANG Ruo-yi;WEI Yong-zhuang;WANG Hui-jiao(Guangxi Key Laboratory of Cryptography and Information Security,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学广西密码学与信息安全重点实验室,广西桂林541004

出　　处：《计算机工程与设计》2023年第6期1629-1634,共6页Computer Engineering and Design

基　　金：国家自然科学基金项目(61872103);广西自然科学基金项目(2019GXNSFGA245004);广西重点研发计划基金项目(桂科AB18281019)。

摘　　要：针对差分神经区分器中准确率随着密码算法轮数增加而快速降低的问题,提出一种差分神经区分器求解方法。将深度学习技术与多差分密码分析相结合,通过采用神经网络拟合密码算法的多输入及多输出差分,设计多差分神经区分器通用模型。该模型中所使用的输入参数被设置为多个明文差分、相应的密文及密文差分。将其应用于分析Speck32/64及Simon32/64密码算法,结果表明,Speck32/64的5至7轮区分器准确率均有显著提升;Simon32/64的密码区分器轮数从9轮提升至10轮,说明该方法的有效性。Aiming at the problem that the accuracy decreases rapidly as the distinguisher round numbers increase,a method for looking for the differential neural distinguisher was proposed by combining both the traditional differential cryptanalysis and the deep learning tools.The multiple differential neural distinguisher model was established using the characteristics of multiple differentials of block cipher algorithm and neural network technique,whose parameters were set as ciphertexts,multiple ciphertext differences and multiple plaintext differences.The method was applied to Speck32/64 and Simon32/64 algorithms.Experimental results show that the accuracies of 5-round,6-round and 7-round Speck32/64 are improved,and the distinguisher of Simon32/64 increases from 9-round to 10-round.All these results illustrate the effectiveness of this approach.

关 键 词：分组密码 差分密码分析 神经区分器 深度学习 多差分 准确率 构建参数 

分 类 号：TP309.7[自动化与计算机技术—计算机系统结构]