The Research Of Cryptosystem Recognition Scheme Based On Machine Learning

In recent year,the amount of network data,which generated by the Internet,increases rapidly in geometric progression.Big data is a huge potential resource for data analyst.However,the network data is usually complexity,diversity and redundant,which have posed a significant challenge for researcher.The prerequisite of encrypted data analysis is to recognize the type of ciphertext.Cryptosystem recognition belongs to distinguish attack,which possess important theoretical significance and applicable value in cryptanalysis.In this paper,we focus on the extraction of ciphertext feature and the construction of machine learning based classifier,and carry out research on their applications in cryptosystem recognition.The main idea and innovation of this paper are as follow:1.The performance of cryptosystem recognition is usually restricted by varieties of conditions.The research on the impacts of recognition is helpful to take the task of cryptosystem recognition.We extracted 14 kinds of ciphertext features and constructed the cryptosystem recognition classifier based on the support vector machine.AES,Blowfish,Camellia,DES and IDEA were taken as recognition targets in experiments.By using the index of pattern recognition,the effects of ciphertext feature,cipher key and the size of file were considered in experiments.The ciphertext feature based on randomness test are used in cryptosystem recognition primarily.2.Enlightened by some randomness test's application in cryptosystem recognition,we firstly proposed 54 kinds of new ciphertext features based on NIST's randomness tests.Combined with Random Forest algorithm,we constructed cryptosystem recognition classifier.Then we completed experiments about the recognition between plaintext and ciphertext,the recognition of ciphertext encrypted by block cipher in ECB mode and CBC mode and the recognition of ciphertexts encrypted by six block ciphers.The effectiveness of features based on randomness tests was verified.Some excellent features were selected through comparative analysis.3.We chose support vector machine,na?ve Bayesian,decision tree,random forest,k-means and k nearest neighbor algorithms to construct the classifier of cipher texts.One to one identification experiments of 5 block ciphers were done to test and verify the efficiency of different classifiers.The performances of different classifiers were analysed.To improve the efficiency of recognition classifier,program parallelization and feature reduction technologies were used to optimize 3 classifiers.The experimental results shown that the run time of Adaboosting and Bagging decreased 61% after optimized by t-SNE algorithm.By using program parallelization technique,the run time of Random Forest decreased 61%.4.The recognition experiments between Grain-128 and other 11 symmetric ciphers were implemented,the experimental results shown the difference between the ciphertexts of Grain-128 and other symmetric cipher.The performance comparison of each features were made.By using t-SNE algorithm,we implemented features reductions to improve six kinds of features' data utilities.With maintaining the stability of the recognition accuracy,the data storage of features decreased 87.5%,and the run time of classification algorithm decreased 95% for best result.