Research On Adversarial Sample Generation Method For Evading Botnet Detection

In recent years,more and more adversarial attack techniques have been proposed.In the fields of computer vision and natural language processing,researchers have explored how to use it to attack existing machine learning or deep learning-based anomaly detection and classification systems,aiming to evaluate the robustness of related technical systems and further promote the improvement of defense methods.However,in the field of Io T security,whether it is a research scholar or a hacker attacker,the research on counterattacks is still in its infancy.Therefore,from the perspective of attack,this thesis supplements the deficiencies of existing research in this regard by deeply exploring the application of adversarial attack technology in Io T security.This thesis first comprehensively analyzes the characteristics of Io T botnets and botnet detection methods based on machine learning or deep learning,and finds that botnet detection systems use different feature selection methods to improve operating efficiency.However,the optimal feature subset obtained by the existing feature selection method is often only applicable to a certain classifier,and the classification accuracy will be greatly reduced when faced with other different classifiers.Therefore,after analyzing the existing mainstream feature selection methods,this paper proposes a new Wrapper-style feature selection method based on FisherScore.This method improves the generation of candidate feature subsets and the evaluation method of candidate feature subsets,and greatly improves the versatility of optimal feature subsets.Experimental results show that on the standard data set UNSW-NB15,the optimal feature subset obtained by this method can make various classifiers obtain an average classification accuracy of 86.80%.Secondly,this paper proposes a new adversarial sample generation method based on a generative adversarial network for black-box attacks on Io T botnet detection systems.This method fits different types of botnet detection systems by training alternative discriminators,and training generators to modify specific statistical characteristics in the original malicious botnet traffic.On the premise of ensuring that its attack characteristics are not changed,it generates adversarial samples that can evade the botnet detection system.Experimental results show that after the malicious network traffic in the standard data set N?Ba Io T is regenerated by this method,the average detection rate of malicious network traffic drops by 73.64%.And it proves that the method is suitable for evading different types of botnet detection systems and botnets composed of different Io T devices,and the generated adversarial samples have good transferability.