Evasion Attack And Its Application In Chinese Spam Message Filtering

Machine learning has been widely used in many security-related applications,such as spam filtering,intrusion detection and malware detection.However,the traditional machine learning technology can't solve the adversarial problems because the traditional machine learning algorithm assumes that the training set and test set should meet the independent and identically distributed hypothesis.In adversarial environment,this assumption does not hold.Therefore,it is necessary to carry on in-depth research on the adversarial environment.This thesis aims at the common evasion attack scenario in adversarial learning,that is,the attacker will modify the features of malicious samples in the test phase to avoid the detection of the classifier and study new evasion attack algorithms.The traditional evasion attack ignores the corresponding constraints for the features of each sample.After adding corresponding constraints,this thesis proposes an evasion attack algorithm based on complex method.It is shown from the simulation experiments on intrusion detection and other datasets that the evasion attack algorithm based on the complex method can not only overcome the defects of the traditional evasion attack algorithm,but also reduce the attack distance and improve the convergence speed of the attack model.In addition,we also study the spam messages filtering in the adversarial environment in this thesis.This thesis mainly considers the form of black box attack,namely the attacker doesn't know the original classifier,and only knows the dataset.Through the good-word attack algorithm,it respectively attacks the classifiers based on the support vector machines,Naive Bayes and logistic regression.Through comparison experiments,it is found that the attack algorithm has migration.Finally,based on the previous method of adversarial machine learning,the anti-spam SMS filtering system in Android environment is designed and developed.The main functions of the system include login interface,registration interface,spam verification,spam interception and notification.