Certified Malware Detection Framework Provides Provable Guarantees Against Adversarial Evasion

arXiv·medium signalXBlueskyLinkedInCopy link

This paper proposes a certifiably robust malware detection framework based on randomized smoothing through feature ablation and targeted noise injection, providing mathematical guarantees against adversarial evasion techniques like metamorphic engine mutations. Unlike standard ML malware detectors that are vulnerable to small perturbations, this approach gives provable bounds on robustness — if a sample is classified as malware, no perturbation within a certified radius can flip the decision.