No Need to Teach New Tricks to Old Malware: Winning an Evasion Challenge with XOR-based Adversarial Samples

Abstract

Adversarial attacks to Machine Learning (ML) models became such a concern that tech companies (Microsoft and CUJO AI’s Vulnerability Research Lab) decided to launch contests to better understand their impact on practice. During the contest’s first edition (2019), participating teams were challenged to bypass three ML models in a white box manner. Our team bypassed all the three of them and reported interesting insights about models’ weaknesses. In the second edition (2020), the challenge evolved to an attack-and-defense model: the teams should either propose defensive models and attack other teams’ models in a black box manner. Despite the difficulty increase, our team was able to bypass all models again. In this paper, we describe our insights for this year’s contest regarding on attacking models, as well defending them from adversarial attacks. In particular, we show how frequency-based models (e.g., TF-IDF) are vulnerable to the addition of dead function imports, and how models based on raw bytes are vulnerable to payload-embedding obfuscation (e.g., XOR and base64 encoding).