Research On Hyperparameter Optimization Algorithm Based On Neural Network Model In Side Channel Attack

As a new way of password attack,the side channel attack is seriously threatening the security of the encryption device.It uses the electromagnetic,energy and other physical information leaked by the encryption device in the process of executing the encryption algorithm to obtain sensitive information.As a learning attack method in the field of side channel attack,template attack is gradually becoming the most powerful attack method in this field.However,the implementation of traditional template attacks requires a lot of human intervention.Therefore,in order to make the template attack process more intelligent,this paper studies the feasibility of template attack based on neural network.At the same time,because the hyper-parameters of neural algorithm affect the final training and attack results to a great extent,so choosing the appropriate hyper-parameters becomes the primary problem of whether to build template attack based on neural network.In order to liberate the staff from the complex and tedious work of parameter adjustment,this paper focuses on how to realize the function of super parameter automatic optimization in the model based on the successful implementation of template attack based on neural network.Firstly,this paper starts with the research status of side channel attack,compares the traditional side channel attack with machine learning algorithm,and reveals the difference and relationship between them.Then,according to the characteristics of traditional template attack and related theories,a side channel attack model based on multi-layer perceptron is established,and the feasibility of the scheme is verified in the experiment based on ASCAD open data set.In order to further improve the intelligence and robustness of the model of side channel attack based on neural network,this paper describes the implementation principle and algorithm flow of three optimization algorithms(Bayesian optimization,grid search,random search).In the experiment,firstly,the template attack based on multi-layer perceptron is compared with the traditional template attack,which proves that the new side channel attack combined with neural network is feasible.Secondly,to realize the automatic optimization of hyper-parameters: in this paper,three different optimization algorithms are applied to the selection of hyper-parameters.The hyperparameters obtained from the optimization algorithm are used to construct the multilayer perceptron,and the constructed multi-layer perceptron is applied to the template attack.In this process,the performance of three optimization algorithms is realized and compared,and the factors influencing the experimental results are analyzed.At the same time,different from the traditional method of dealing with the discrete value in Bayesian optimization,a method of optimizing the parameters of side channel multilayer perceptron is developed,which can combine the experience of hyper-parameters,and further improves the efficiency of the optimization algorithm.A series of experiments show that the choice of hyper-parameters has a great influence on the experimental results.At the same time,the corresponding multi-layer perceptron model for template attack is constructed by different automatic optimization methods.After that,in order to analyze and compare the characteristics and efficiency of the three different optimization algorithms,this paper compares the eight hyperparameters required by MLP from two aspects of recommended times and time cost.The experimental results show that:(1)the feasibility of using automatic optimization algorithm to build multi-layer perceptron suitable for side channel attack is verified;(2)the influence of different optimization methods on the experimental results under different recommended times is compared,and the performance of building model based on artificial optimization and automatic optimization method in actual attack is compared;(3)the performance of different optimization methods in actual attack is known Finally,based on the characteristics of model building,the Bayesian optimization algorithm is determined as the best optimization algorithm to build the template attack model based on multi-layer device.