FAME: From Adversarial Feature Vectors to Malware Evasion via SMT

Abstract

Internet of Things (IoT) malware detection remains a challenging task due to the severe resource constraints of IoT devices, which require lightweight yet effective security solutions. SIMBIoTA-ML addresses this challenge by leveraging machine learning (ML) classifiers operating on TLSH-based static features for efficient malware detection. However, like other ML-based systems, SIMBIoTA-ML is vulnerable to adversarial evasion attacks. While prior work investigated this problem through ad-hoc binary modifications, it did not provide a systematic or guaranteed approach for generating adversarial examples. In this paper, we propose FAME (From Adversarial feature vectors to Malware Evasion), a framework for generating adversarial malware samples against SIMBIoTA-ML. FAME applies established ML evasion attacks in the feature space and addresses the inverse feature-mapping problem between continuous adversarial feature vectors and corresponding executable binaries. To this end, we formulate the discretization of continuous adversarial feature vectors as a Satisfiability Modulo Theories (SMT) problem, enforcing both TLSH validity constraints and SIMBIoTA-ML evasion constraints. Using the target TLSH hash determined by the SMT solver, the original malware binary is then modified accordingly. By combining adversarial ML, constraint solving, and controlled binary modification, FAME enables the scalable and reliable generation of executable adversarial malware samples.