摘要
Model stealing (模型窃取 MS) involves(涉及) querying and observing the output of a machine learning model(查询和观察机器学习模型的输出) to steal its capabilities(窃取其功能). The quality of queried data is crucial(查询数据的质量是至关重要的), yet obtaining a large amount of real data for MS is often challenging(然而获取大量的真实数据通常是具有挑战性的). Recent works have reduced reliance on(减少依赖) real data by using generative models(生成模型). However, when high-dimensional query data(高维查询数据) is required, these methods are impractical(不切实际) due to the high costs of querying(查询成本高) and the risk of model collapse(模型奔溃的风险). In this work, we propose using sample gradients(样本梯度 (SG) to enhance the utility of each real sample(增强每个真实样本的效用), as SG provides crucial guidance(至关重要的指导) on the decision boundaries(决策边界) of the vic