An Adversarial Machine Learning Method Based on OpCode N-grams Feature in Malware Detection

Machine learning has become an important method in malware detection. However, due to the weakness of machine learning models, a large number of researches related to adversarial machine learning has emerged. At present, the researches about adversarial machine learning mainly focus on the image and speech recognition. In the field of malware detection, because the feature modification can easily damage the integrity and functionality of the code, it is usually through adding noise such as garbage instructions to achieve fooling the malware detection model. In this paper, we propose an adversarial machine learning method on malware detection model based on OpCode n-grams feature. We first collect a large number of malicious code and normal code data sets, and use TF-IDF to extract OpCode n-grams with different n values from the data set. Then we train three malware detection models based on OpCode n-grams. Under the premise of comprehensive consideration of efficiency, accuracy and interpretability, we select XGBoost as the adversarial feature extraction model and extract adversarial features. Finally, in order to verify the accuracy of the extracted features, we conduct an adversarial machine learning experiment. The experimental results show that the adversarial method proposed in this paper can completely fool the machine learning detection model.