Adversarial Attacks in Problem Space for VBA Code Samples

Machine learning is being used more frequently in cyber security. One area is in antivirus software. Learning-based antivirus software is not perfect, since threat actors can use adversarial data to craft attacks against these learning-based security products. This paper focuses on using machine learning to craft attacks against Microsoft (MS) Windows Defender. The goal of this work is to provide tooling and knowledge for: security vendors to tighten the security of their products, security professionals to aid in showcasing risks to customers, and law enforcement to help discover the locations and identities of cyber criminals. The attacks focus on modifying Visual Basic for Applications (VBA) code samples to flip Defender’s classification of those files. The modifications are done to the code itself (i.e., problem space) rather than just in feature space. The project attempts to extract information from MS Windows Defender to help craft the attacks. Preliminary results showed that the classification of some of the files was changed after targeted modification and that the attacks were feasible. Additional work will be conducted to further increase the reliability of the attacks.