Using Adversarial Examples Enhancing Advanced Persistent Threat Detection

Advanced persistent threats are long-term network attacks on specific targets with attackers using advanced attack methods.In the field of advanced persistent threat detection,most works used machine learning methods whether host-based detection or network-based detection.The host-based detection methods are mainly to detect whether there are malicious behaviours on independent hosts such as the execution of malicious software,the behaviour of applications trying to modify certain files.The network-based detection methods usually take network flow data as input and aim to find abnormal network packets and abnormal network interactions through statistical analysis,data mining or machine learning.Both host-based detection methods and network-based detection methods have adopted machine learning methods mostly.In recent years,great progress has been made in using machine learning methods to detect APT attacks.However,the current work has the following problems:1.Machine learning methods lack robustness because they are vulnerable to adversarial examples.Machine learning models are susceptible to adversarial examples due to their over-linearization,adversarial examples may affect the APT detection model so that it cannot effectively detect APT.2.There is no method to defend against adversarial attacks on the APT detection model or to increase the robustness of the APT detection model.Existing APT detection models are unaware that they will be attacked by adversarial examples.This paper mainly studies the above problems and the specific work is as follows:1.This paper attacks APT detection models with adversarial examples.This paper first trained four APT detection models with different machine learning methods and highest F1-score of models is 0.9791.According to the characteristic of APT traffic,this paper proposed the adversarial examples generation algorithm for the APT detection model.In this paper,gray-box and black-box attacks for the APT detection model are carried out through algorithm proposed.The gray-box attack reduced the detection success rate of the SVM model from 98.52% to 1.47%.The detection success rate of the model with the best black-box attack effect decreased from 78.26% to 0.13%.The experimental results show that the machine learning-based APT attack detection model can be successfully attacked with adversarial samples.2.This paper adversarial train the APT detection model with adversarial examples.The experimental results show that after the adversarial training,the success rates of the SVM model,the random forest model and the logistic regression model have increased to more than 99%.That is,the adversarial training improves the robustness of these model...