Design And Analysis Of Lightweight Stream Cipher Based On Machine Learning

With the rapid development of computer and communication technology,more and more data are transmitted on the public Internet.In order to protect data security,it is necessary to understand the means of the attacker.On this basis,the design of encryption algorithms that can enhance the ability to resist attacks is absolutely necessary.In cryptographic attacks,cryptanalysts can only intercept ciphertexts encrypted by unknown algorithms.As a result,the work of cryptanalytic decoding cannot be carried out.Therefore,the identification of cryptanalytic algorithms is the first step in cryptanalysis.Machine learning has high insight and optimization capabilities for data processing.Among them,support vector machines and random forests are often used to classify data and have attracted attention in the field of cryptosystem recognition.Secondly,the structure of machine learning algorithms is highly non-linear.Designing cryptographic algorithms can increase the difficulty of cryptanalysis and provide a basis for the design of machine learning in cryptographic structures.At present,the study of machine learning methods in stream cipher has just begun.With less research results,the design and analysis of lightweight stream cipher is relatively rare.Meanwhile,the hardware implementation of neural networks has parallelization and fast processing speed.It's easier implementation of lightweight stream cipher in hardware.This thesis uses machine learning methods to study the identification analysis and design of lightweight stream cryptosystems.The main research is divided into the following two aspects:First,based on the research results of cryptosystem recognition of block ciphers,the recognition model and related indicators of light-weight stream cipher systems are given.For the first time,five lightweight stream ciphers of Fruit-80,Sprout,Plantlet,Grain,and Lizard are tried to identified.The algorithms are implemented separately and the actual plaintext is encrypted to obtain the ciphertext file.Then,randomness detection,fixed-length bits,and entropy are used to extract ciphertext features to construct a data set.Finally,in different task scenarios,support vector machines and random forest recognition schemes are used for classification training and testing.The following results are obtained through experiments: when using the support vector machine algorithm,the recognition accuracy of the two classifications reaches about 60%,and 30.96% in the multi-classification.Both are better than random guessing.When the random forest algorithm is used,the recognition accuracy of the two classifications is more than 85%,and about 55% for multi-classification.When the parameters of the two classification algorithms are reasonable,the recognition efficiency of the random forest algorithm is higher than that of the support vector machine.Second,BP neural network is an extension of multilayer perceptron,which solves the problem of perceptron network connection weight learning.Therefore,BP neural network is more conducive to the realization of Boolean functions.This thesis analyzes the feasibility of implementing Boolean functions in BP neural network based on the existing quaternion Boolean functions that are easy to implement using perceptron.Furthermore,it was found that the Boolean function can be realized by using the over-fitting phenomenon that occurs during the learning process of the BP neural network.A Boolean function higher than quaternion is obtained through an example analysis.It can be known that this way of implementing Boolean functions is better than the existing perceptron methods,and higher-order Boolean functions can be obtained.In order to enhance the security of the Fruit-80 algorithm.The 25 p network based on multilayer perceptron can realize two n-1 bits Boolean functions to form n bits Boolean functions.Then the nonlinear module of the Fruit-80 algorithm is designed by using 25 p perceptron and BP neural network.This thesis lists the specific structure and initialization process of the algorithm.After analysis,the scheme can learn different networks(different nonlinear modules)according to different users' key training.It is proved that Fruit-80 algorithm based on BP neural network can enhance resistance to TMDTO attack and algebraic attack.At the same time,because the neural network has advantages in hardware implementation,the algorithm has good use value.Finally,NIST and national secret pseudo-random detection are performed on the sequence generated by the algorithm.