MACHINE LEARNING TECHNIQUES TO PERCEIVE INTRUSIVE SOFTWARE

Abstract

Machine learning based malware detection algorithms are usually integrated into antivirus software and therefore they are black-box systems to malware authors. It is hard for malware authors to know which classifier a malware detection system uses and the parameters of the classifier. However, it is possible to figure out what features a malware detection algorithm uses by feeding some carefully designed test cases to the black-box algorithm. Malware authors can modify a benign program’s DLL or API names to malware’s DLL or API names, and vice versa. If the detection results change after most of the modifications, they can judge that the malware detection algorithm uses DLL or API features. Therefore, in this paper we assume that malware authors can know what features a malware detection algorithm uses but know nothing about the machine learning model. The proposed model in this paper uses GANs, which generates adversarial examples to attack black-box malware detection algorithms. A substitute detector is trained to fit the black-box malware detection algorithm, and a generative network is used to transform malware samples into adversarial examples. Experimental results show that almost all the adversarial examples generated by GANs successfully bypass the detection algorithms.