Deep Learning Analysis Of Hardware Cryptographic Chip Side Channel Residual Power Consumption

Cryptography is an important foundation for modern information security,yet the security of cryptographic algorithms by design does not guarantee that they will be equally secure when implemented in hardware.The implementation of cryptographic algorithms on cryptographic hardware devices is accompanied by the leakage of physical information such as time,power consumption and electromagnetic radiation,which is used by attackers to recover the secret key through side-channel attacks.Among the side channel attacks,the power consumption attack has developed the most rapidly and has become a powerful method of secret key attack on many cryptographic chips.Therefore,it has become an important task for the information security industry to study attacks on cryptographic chips and discover cryptographic chip security vulnerabilities in advance.In this paper,we study side channel attacks on cryptographic chips based on AES-128 encryption algorithm,and the main work and research results are as follows:Firstly,the principle of side channel attack and the traditional side channel attack scheme are introduced.Compared with the traditional side channel attack method,the deep learning based side channel attack greatly improves the attack efficiency.The paper focuses on the principle and attack steps of the deep learning-based side channel attack method,and based on this,the study of the residual power consumption traces is carried out.A hardware encryption chip power acquisition and analysis experimental setup was established using the chipwhisperer data acquisition platform to collect the power consumption leakage of AES-128 encryption algorithm when running in the hardware chip and to pre-process the data.A side-channel attack based on the Long Short Term Memory network（LSTM）was proposed.The LSTM-based side channel analysis method is experimentally compared with two other commonly used deep learning side channel analysis methods,i.e.,with the side channel analysis based on Multi Layer Perceptron（fully connected）neural network（MLP）,Convolutional Neural Network（CNN）-based The experimental results show that the long and short term memory network proposed in this paper has the highest accuracy rate in single secret key byte attack,and the long and short term memory network has higher attack efficiency than the other two networks in recovering the complete secret key,and can be more effective in vulnerability detection.This paper shows that the proposed long and short term memory network has the highest accuracy in single secret key byte attacks.In addition,the actual attack analysis may have incomplete energy trace collection,in order to improve the utilization of energy traces,based on the study of the complete power consumption traces,this paper investigates the actual incomplete collection of crippled power consumption traces that often occur in practice.Firstly,the reasons for the emergence of the residual power consumption are analysed,then the characteristics of the residual power traces are investigated,the leakage interval is explored in detail,a method for setting residual power labels is proposed,and an incomplete power consumption dataset is produced;then by introducing the Connectionist Temporal Classification algorithm（CTC）algorithm,combined with LSTM network and Bidirectional Long Short Term Memory network（BLSTM）,to build a CTC-BLSTM attack model;finally,for the crippled power consumption traces in different cases,we use the input conventional single-byte attack model（MLP,CNN,LSTM attack model）and CTC-BLSTM model to attack the crippled power consumption attack respectively,the experimental results show that:（1）when the crippled power consumption trace length is less than or equal to the single-byte leakage interval,the accuracy rates of both are similar;（2）when the energy trace is complete data,the accuracy rate of the single-byte attack model is slightly higher than that of the CTC-BLSTM model;（3）when the power trace length is larger than the complete power consumption and smaller than the single-byte leakage interval,the CTC-BLSTM model attack accuracy is much higher than that of the single-byte model.The above three cases,in any case CTC-BLSTM model attack energy traces only need a model to be able to achieve a high accuracy rate,using the mean-models algorithm proposed in this paper can easily obtain the real secret key sequence,while single-byte attack model in favorable circumstances（（1）（2））still need 16 models to obtain the real secret key sequence,through the above experimental results can be seen The CTC-BLSTM attack model achieves better attack efficiency in both complete power traces and crippled power traces analysis.