【论文阅读】AugSteal: Advancing Model Steal With Data Augmentation in Active Learning Frameworks（2024）

摘要

With the proliferation of（随着） machine learning models（机器学习模型） in diverse applications, the issue of model security（模型的安全问题） has increasingly become a focal point（日益成为人们关注的焦点）. Model steal attacks（模型窃取攻击） can cause significant financial losses（重大的经济损失） to model owners（模型所有者） and potentially threaten（可能威胁） the security of their application scenarios（程序场景的安全性）. Traditional model steal attacks（传统模型窃取攻击） are primarily directed（主要针对） at soft-label black boxes（软标签黑盒）, but their effectiveness significantly diminishes（有效性大大降低） or even fails in（甚至失效） hard-label scenarios（硬标签场景）. To address this, for hard-label black boxes（对于硬标签黑盒）, this st