Deceiving Portable Executable Malware Classifiers into Targeted Misclassification with Practical Adversarial Examples

Due to voluminous malware attacks in the cyberspace, machine learning has become popular for automating malware detection and classification. In this work we play devil's advocate by investigating a new type of threats aimed at deceiving multi-class Portable Executable (PE) malware classifiers into targeted misclassification with practical adversarial samples. Using a malware dataset with tens of thousands of samples, we construct three types of PE malware classifiers, the first one based on frequencies of opcodes in the disassembled malware code (opcode classifier), the second one the list of API functions imported by each PE sample (API classifier), and the third one the list of system calls observed in dynamic execution (system call classifier). We develop a genetic algorithm augmented with different support functions to deceive these classifiers into misclassifying a PE sample into any target family. Using an Rbot malware sample whose source code is publicly available, we are able to create practical adversarial samples that can deceive the opcode classifier into targeted misclassification with a successful rate of 75%, the API classifier with a successful rate of 83.3%, and the system call classifier with a successful rate of 91.7%.