一句话：以 ClamAV 为白盒参照，提出 RAVEN 测量反病毒引擎在语义保持扰动下的检测不稳定性。
论文指出现实反病毒引擎仍大量依赖基于特征码的检测，而学术对抗研究主要面向机器学习模型，造成评估空白。作者以 ClamAV 的开源特征库为白盒参照，分析其检测机制，发现原始字节级特征占比达 94.54%。在此基础上，论文提出 RAVEN，通过语义保持扰动来测量 AV 引擎检测结果的不稳定性，即在不改变程序语义的前提下观察检测是否发生波动。评估围绕特征库分析与扰动实验展开，旨在揭示特征码检测的脆弱性，并为更稳健的 AV 评估提供方法。
Real-world anti-virus (AV) engines continue to rely heavily on signature-based detection, whereas academic adversarial research has largely focused on ML-based models, leaving an evaluation gap. To better understand signature-based detection behavior, we analyze ClamAV’s open-source signature database as a white-box reference point. Our analysis shows that raw byte-level signatures account for 94.54
@inproceedings{paperbot12450,
title = {RAVEN: Measuring Detection Instability in Anti-virus Engines under Semantics-Preserving Perturbations},
author = {Younghoon Ban and Beomjin Jin and Doowon Kim and Hyoungshick Kim and Haehyun Cho},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3837492}
}