DOMAIN KNOWLEDGE-DRIVEN ADVERSARIAL EXAMPLE GENERATION METHOD FOR ANDROID SOFTWARE

Android software adversarial example is a research hotspot in the field of mobile malware detection. The existing generation methods only consider adversarial example generation from the perspective of gradient and algorithm, while ignoring the inherent characteristics of Android software, resulting in problems such as low example generation efficiency and unreliable example quality. This paper proposes a domain knowledge-driven adversarial example generation method named DK-JSMA for Android software. It used text analysis, code analysis, and expert knowledge to extract and quantify Android domain knowledge. It used the above domain knowledge to optimize the selection strategy and selection result of the feature perturbation process. The experimental results on large-scale data sets show that the DK-JSMA method has higher generation efficiency, and compared with the classic JSMA method, it reduces the feature distortion by over 20%, 23% and 44% than DNN, SVM, and RF.