Research On Robustness Testing Method Of Malicious DNS Traffic Detection Model

With the continuous development of the Internet,network security problems are becoming more and more prominent.Currently,malicious traffic based on DNS is gradually increasing,and most intrusion detection systems release DNS traffic directly by default.The resulting information leakage problem will bring great security threats to network users.Some scholars try to use machine learning and other methods to detect DNS malicious traffic,so as to improve the security of network and information systems.At present,many advanced DNS traffic network intrusion detection systems are equipped with machine learning modules to increase detection ability.The detection model can achieve good traffic classification results.However,the machine learning model itself has the problem of weak robustness,which is easy to be disturbed by adversarial samples,which can reduce the detection performance of intrusion detection systems and bring huge security threats.Therefore,this research mainly uses adversarial samples to test the robustness of DNS traffic intrusion detection systems based on machine learning.At present,research on adversarial samples is generally concentrated in the fields of images.Most existing research in the field of network traffic focuses on the extracted features rather than the traffic data itself.At the same time,most of them assume that the attacker has a certain understanding of the target model in advance,which does not meet the needs of the real network scene and cannot be put into actual evaluation and testing,Therefore,this research aims to solve the above difficulties and test the robustness of the malicious DNS traffic detection model.The main research contents and results of this research are as follows:1.Design and implement the generation scheme of malicious DNS traffic adversarial samples.The scheme includes two parts: adversarial feature generation and traffic packet generation.The adversarial feature generation part realizes adding noise in the feature space,and the traffic packet generation part completes the generation of adversarial samples in the traffic space by using the three dimensions of packet length,number of packets and time.The adversarial samples constructed in this way can meet the constraints of network protocol and have the function of network circulation.2.Design the robustness test scheme of malicious traffic detection model in the white-box scenario and black box scenario.Different machine learning algorithms are used to build the target model.According to the different understanding and query restrictions of the target model in the white-box scenario and black-box scenario,the robustness test framework and process are designed respectively.Combined with the malicious DNS traffic adversarial sample generation scheme in this research,the robustness of the target model is tested directly in the white-box scenario.The alternative model is built through model extraction in the black-box scenario to complete the generation of countermeasure samples.Then the robustness of the target model is tested by using adversarial samples.The results show that the robustness of the machine learning model will be affected by the adversarial samples,and the robustness of the model built with the same algorithm in the black-box scenario is stronger than that in the white-box scenario,which solves the problem that it is difficult to generate the adversarial samples in the black-box scenario.3.Verify the reliability of the proposed test scheme.In this research,a comparative experiment is designed to test the robustness of the target model after adversarial training.The results show that the robustness of the model is enhanced after adversarial training.It also shows that the scheme proposed in this research is reliable in testing the robustness of malicious DNS traffic detection model,and can provide conditions for promoting the development and improvement of intrusion detection systems.