【论文阅读】Enhance Model Stealing Attack via Label Refining（2022）

摘要

With machine learning models（机器学习模型） being increasingly（越来越多） deployed（部署）, model stealing attacks（模型窃取攻击） have raised an increasing interest. Extracting decision-based models（基于决策的模型窃取） is a more challenging task with the information of class similarity missing（类相似度信息缺失的情况下）. In this paper, we propose a novel（新颖） and effective（有效） model stealing method as Label Refining via Feature Distance (LRFD 基于特征距离的标签精炼), to re-dig the class similarity（重新挖掘类相似度）. Specifically（具体来说）, since the information of class similarity（类相似度信息） can be represented by the distance between samples（样本之间的距离） from different classes（不同类） in the feature space（特征空间）, we design a soft label（软标签） construction module（构造模型） inspired by the prototype learning（原型学习）, and transfer the knowledge in the soft label to the substitution model（替代模型）. Extensive experiments conducted on（进行了大量的实验） four widely-used（广泛使用的） datasets consistently demonstrate（一致表明） that our method yields（产生） a model with significantly greater functional similar