Faculty Mentor Information
Dr. Jyh-Haw Yeh (Mentor), Boise State University
Additional Funding Sources
Supported by National Science Foundation Award #2244596 REU Site: Summer 2024 Cloud Computing Security and Privacy.
Abstract
Malicious software, commonly known as malware, refers to any type of intrusive software designed to perform harmful actions on a computer system. Recently, Machine Learning (ML) techniques have been used to create new malware variants, enabling attackers to generate thousands of previously unseen malware samples. Traditional detection methods, such as signature-based detection, rely on prior knowledge of malware and therefore often fail to identify new variants. This limitation has led cybersecurity experts to increasingly adopt ML techniques for malware detection.
While ML-based approaches have shown promising results by generalizing malware signatures to detect previously unseen malware, they remain vulnerable to adversarial attacks. Adversarial attacks leverage carefully crafted malware samples designed to evade ML-based detectors by exploiting algorithmic vulnerabilities. To develop new defense methods against these attacks, a clear understanding of adversarial techniques is essential.
This study compiles and categorizes the latest research on adversarial attacks in the field to support researchers in developing robust malware detection models. It expands on existing surveys by analyzing adversarial attacks based on attack settings, techniques, success rates, evaluation metrics, and future research directions. This study also proposes promising areas for future research, aiming to highlight gaps in the current body of knowledge.
Survey on Adversarial Attack for Malware Detection
Malicious software, commonly known as malware, refers to any type of intrusive software designed to perform harmful actions on a computer system. Recently, Machine Learning (ML) techniques have been used to create new malware variants, enabling attackers to generate thousands of previously unseen malware samples. Traditional detection methods, such as signature-based detection, rely on prior knowledge of malware and therefore often fail to identify new variants. This limitation has led cybersecurity experts to increasingly adopt ML techniques for malware detection.
While ML-based approaches have shown promising results by generalizing malware signatures to detect previously unseen malware, they remain vulnerable to adversarial attacks. Adversarial attacks leverage carefully crafted malware samples designed to evade ML-based detectors by exploiting algorithmic vulnerabilities. To develop new defense methods against these attacks, a clear understanding of adversarial techniques is essential.
This study compiles and categorizes the latest research on adversarial attacks in the field to support researchers in developing robust malware detection models. It expands on existing surveys by analyzing adversarial attacks based on attack settings, techniques, success rates, evaluation metrics, and future research directions. This study also proposes promising areas for future research, aiming to highlight gaps in the current body of knowledge.