Manually constructed malware analysis platforms that identify important capabilities in malicious software cannot keep up with the massive amounts of malware being released on a daily basis. Traditional approaches that detect the functional capabilities of malware usually contain brittle handcrafted heuristics that quickly become outdated, and can be exploited by nefarious actors. As a result, it is necessary to change the way software security is approached by using advanced analytics, i.e., machine learning, and significantly more automation, to develop more adaptable malware analysis engines that correctly deduce the important capabilities of malware. In this talk, we discuss using machine learning to find accurate models for the prediction of malware capabilities. We explore the characterization of unseen malware binaries using features extracted from extremely fast static analysis of malicious code, as well as information derived from slower dynamic analysis that comes from the execution of binaries in a malware analysis sandbox. Our experimental results demonstrate that by learning from relatively large amounts of malware, we are able to accurately predict important capabilities of malicious executables with an accuracy of up to 97.70%.
Audience: Anyone interested in applications of AI for cybersecurity