Opcode Sequence Amplifier using Sequence Generative Adversarial Networks

Changhee Choi, Sunguk Shin, Inseop Lee

Abstract: Weapon systems and military networks are threatened by increased cyber attacks. In case of usual cyber attack, commercial grade defence systems are available. Because of speciality of military system, cyber attacks for them are also special and the number is extremely small. For machine learning based defence system, the shortage of malware samples is critical problem. To solve this problem, we proposed the new method for amplifying opcode sequence which is part of the malware and used for malware detection. We first extract opcode sequences from malwares and benign portable files. To make them meaningful and easy to learn, whole opcode sequence is split into several blocks, called OPSEN(OPcode SENtence), using special delimiters. Considering that opcode sequence is not a numerical data but a sequence of instruction, we used SeqGAN with stochastic policy in reinforcement learning and policy gradient. The experimental results shows that the proposed amplified opcode sequence help to improve the detection rate.