Research And Implementation Of Side Channel Analysis Method Based On Machine Learning

With the development of the Internet,people gradually realize that it is very important to protect the security of network devices with password function.In daily life,password devices are widely used,such as campus cards,bank cards,access control cards,and even high-end weapons and equipment with integrated password function.The encryption function is the core of the cipher function.In general,the decryption process of encryption functions is either computationally expensive or time-consuming.Encryption algorithm is difficult to crack in mathematics theory to ensure the security of cryptographic equipment.However,side channel analysis theory obtains confidential information from physical leaks rather than mathematical algorithmic theoretical flaws.Compared with traditional methods,side channel analysis takes less time and is more efficient in cracking encryption algorithms.This thesis analyzes the research status of machine learning in side channel analysis.Aiming at the Advanced Encryption Standard(AES)of symmetric Encryption algorithm,a mask AES side channel analysis model based on machine learning is designed to improve the efficiency of side channel attack.The main research contents of this thesis are as follows:(1)A side channel analysis model based on SAE was proposed to study the leakage information of embedded FPGA cryptochip equipment during operation.In this thesis,a variety of feature extraction methods are explored through several experiments.The experiment shows that Pearson correlation coefficient method is the most effective method among the explored methods.Aiming at the problem of low efficiency of classifier caused by class imbalance in feature data set,this thesis adopts various data set enhancement methods,such as random oversampling,synthetic minority oversampling algorithm and three variants of synthetic minority oversampling algorithm,which can effectively improve the classification efficiency.In view of the problem that mask protection hinders decrypting key in encryption algorithm,this thesis adopts Naive Bayes,KNN,SAE and discriminant analysis classification method to decrypt mask.This experiment shows that discriminant analysis method is the best attack method.In order to solve the key cracking problem,this thesis proposes a side channel attack model based on S AE,and then compares the model with the NN model for experimental analysis.Experimental results show that the proposed model improves the attack success rate by about 10%compared with NN model.However,in the experimental process,each layer of SAE constructed needs more iterations,and the construction of multi-layer neural network requires a long training time.(2)In view of the deep neural network multi-layer architecture,the number of iterations is large,the neural network training time is long,leading to the side channel attack time-consuming,a side channel analysis model based on integrated learning is proposed.The influence of the combination of different strong and weak classifiers on the side channel analysis is explored,the ensemble learning method with the best classification effect is explored,and the influence of the number of characteristic samples in the data set on the experimental results is explored.In order to solve the mask cracking problem,four machine learning methods are tried in this experiment,including decision tree,random forest,BP algorithm and Adaptive Boosting algorithm.The experiment shows that the Ada-Boost method can restore the mask effectively.For key cracking problems,put forward a kind of Ada-Boost side channel analysis method based on linear discriminant analysis,by comparing the attack results of Ada-Boost method,SAE and NN,the experiment shows that the proposed Ada-Boost model has the highest attack success rate,and the average training time of the model is reduced,and the efficiency is greatly improved.