基于PU分类的差分区分器及其应用

Differential Distinguisher Based on PU Learning and Its Application

差分分析方法的核心是构造高效的差分区分器.2019年Aron Gohr采用深度学习残差网络的方法构造差分区分器,应用于减轮Speck32/64密码算法,五轮和六轮的差分器成功率分别是0.929和0.788.本文采用PU学习(positive-unlabeled learning)的方法,对Speck32/64算法的差分对数据进行训练,利用神经网络中的多层感知机与基于PU学习构造的损失函数,训练得到了一个基于PU分类的差分区分器,并对于减轮Speck32/64算法进行攻击,五轮和六轮差分器成功率分别是0.965和0.860.

The core of differential attack is to construct an efficient differential distinguisher.In 2019,Aron Gohr proposed a deep learning residual network method to construct a differential distinguisher.For the reduced-round Speck32/64 cryptographic algorithm,the success rates of the five-round and six round distinguisher are 0.929 and 0.788,respectively.This paper uses the multi-layer perceptron and the loss function based on PU-learning(positive-unlabeled learning)to train the differential pair data of the Speck32/64 algorithm to obtain a differential distinguisher based on PU learning.Based on this distinguisher,for the reduced round Speck32/64 algorithm,the success rates of the five rounds and six rounds are 0.965 and 0.860,respectively.