Abstract

AI drastically reduces the effort required to produce malware that mutates both its code and behavior, thereby creating polymorphic and nondeterministic variants in which traditional signatures and many heuristic defenses fail. In this paper, I survey recent developments in AI-assisted malware generation, explain why conventional defenses are insufficient, and propose a layered detection architecture emphasizing semantic behavior, streaming anomaly detection, and defensive generative augmentation. I also outline why this approach generalizes to previously unseen AI-mutated samples, provide an evaluation plan with meaningful metrics, and describe a feasible MVP roadmap for practical deployment. Recent disclosures and threat intelligence further highlight the urgent need for semantic-focused detection rather than purely syntactic checks.

Faculty Advisor/Mentor

Yue Xiao

Document Type

Paper

Disciplines

Computer Sciences | Cybersecurity

DOI

10.25776/dtg4-4v41

Publication Date

11-10-2025

Upload File

wf_yes

Included in

Cybersecurity Commons

Share

COinS
 

Detecting Generative-AI-Enabled Polymorphic Malware: A Semantic-Behavior Approach

AI drastically reduces the effort required to produce malware that mutates both its code and behavior, thereby creating polymorphic and nondeterministic variants in which traditional signatures and many heuristic defenses fail. In this paper, I survey recent developments in AI-assisted malware generation, explain why conventional defenses are insufficient, and propose a layered detection architecture emphasizing semantic behavior, streaming anomaly detection, and defensive generative augmentation. I also outline why this approach generalizes to previously unseen AI-mutated samples, provide an evaluation plan with meaningful metrics, and describe a feasible MVP roadmap for practical deployment. Recent disclosures and threat intelligence further highlight the urgent need for semantic-focused detection rather than purely syntactic checks.