Research And Application Of Deep Learning In Side Channel Analysis

Side channel analysis(SCA)is an efficient and easy to implement method in the security analysis of cipher chips,which poses a great threat to the chip security.Therefore,the international major evaluation organizations regard the ability to resist the SCA as one of important evaluation metric to investigate the chip security.Cipher chip is widely used in modern life,its security is directly related to people's livelihood,medical,military and so on.To study more efficient side channel analysis method is helpful to evaluate the security of cipher and provide reference for the design of cipher algorithm.In recent years,with the rapid development of big data and artificial intelligence,more and more researchers begin to study the new analysis scheme combining side channel analysis with machine learning and deep learning.In this paper,a side channel classification model based on auto-encoder and a side channel attack inversion model are proposed,which combines deep learning and traditional modeling side channel attack and non-profiling side channel attack.Finally,a mask scheme for anti-SCA is proposed.The main work of this paper is as follows:(1)This paper analyzes the principle of multi-layer perceptron and convolutional network in deep learning,and proposes a classification model based on autoencoder.Firstly,the side channel information such as power consumption,electromagnetic and so on is encoded.Then,a deep convolutional classification network is used to obtain the classification probability of the key from the power consumption samples.In the experiment,based on the dataset provided by ASCAD,the model is trained and verified.The results show that the key can be recovered correctly in the case of 120-150 samples,and the accuracy has been greatly improved.(2)After the simulation of power consumption in non-profiling side channel attack is transformed into an inversion model,and a theoretical model from power consumption to correlation median is established.In this paper,we consider the characteristics of power consumption related to intermediate value in non-profiling attacks.In the model design,we extract and combine the features of each layer of power consumption samples to get the inverse value of intermediate value.In function,it solves the problem that only one key byte can be recovered at the same time in traditional side channel attack.Two data sets are used in the model training,one is the unprocessed power consumption data in ASCAD,the other is the power consumption sample data collected by FPGA implementation of AES and present on Sakura-G platform.The experimental results show that the correlation of the model on ASCAD data set is above 0.8,and the evaluation indexes of AES and PRESENT algorithm in FPGA are 0.95.In terms of attack efficiency,only 20-30 samples are needed to recover the key.The inversion model greatly improves the applicability and accuracy.(3)Finally,this paper discusses the side channel analysis defense scheme under the background of deep learning.The AES algorithm is studied and reconstructed and a high-order mask scheme based on zero knowledge proof is proposed.This scheme randomly divides the sensitive median into d+1 parts,and each d part is independent of the median,so as to achieve the hiding of the median.The experimental results show that this scheme can effectively hide the median It can resist all kinds of side channel analysis and provide effective protection for the security of cryptographic devices.