【论文阅读】DivTheft: An Ensemble Model Stealing Attack by Divide-and-Conquer（2023）

摘要

Recently, model stealing attacks（模型窃取攻击） are widely studied（广泛研究） but most of them are focused on stealing a single non-discrete model（窃取单个非离散模型）, e.g., neural networks（神经网络）. For ensemble models（对于集成模型）, these attacks are either non-executable（不可执行） or suffer from intolerant performance degradation（遭受不可容忍的性能损失） due to（由于） the complex model structure（复杂网络结构） (multiple sub-models 多个子模型) and the discreteness possessed by the sub-model（子模型拥有的离散性） (e.g., decision trees 决策树). To overcome the bottleneck（为了克服这一瓶颈）, this paper proposes a divide-and-conquer strategy（分而治之的策略） called DivTheft to formulate the model stealing attack（模型窃取攻击） to common ensemble models（常见的集成模型） by combining active learning（结合主动学习） (AL). Specifically（具体来说）, based on the boosting learning concept（基于增强学习的概念）, we divide（划分） a hard ensemble model stealing task（硬集成模型窃取任务） into multiple sim