【摘要】 [目的/意义]在模型提取攻击中，攻击者使用的数据极大地影响了攻击的有效性。在实践中，攻击者往往很难获取到精确的目标数据。相比之下，目标模型所用数据的分布可能更容易获得。基于目标分布进行模型提取攻击的方法，为攻击提供了新的思路和方法，同时也提醒拥有有价值模型的云平台和个人，需要采取措施防止攻击者获取目标分布。[方法/过程]使用一些生成对抗网络来学习目标分布并生成数据，并使用此数据执行模型提取攻击。研究了提取部分类攻击方法，并提供了两种提取部分类的方法，丰富了模型提取攻击的攻击手段。[结果/结论]实验结果表明，基于目标分布进行模型提取是可行的，得到的替换模型甚至在某些类别上超过了目标模型的性能，同时攻击效果随着使用的分布接近目标分布而增强。更多还原

【Abstract】 [Purpose/Significance] In model extraction attacks, the data used by the attacker greatly affects the effectiveness of the attack. In practice, it is often difficult for an attacker to obtain the exact target data.In contrast, the distribution of the data used in the target model may be more readily available. This paper investigates a method for model extraction attacks based on target distributions. This provides new ideas and methods for attacks and also reminds cloud platforms and individuals with valuable models of the need to take steps to prevent attackers from obtaining target distributions.[Method/Process] This paper uses several generative adversarial networks to learn the target distribution and generate data and uses this data to perform model extraction attacks. This paper also investigates the method of extracting partial class attacks and provides two methods for extracting partial classes. This enriches the attack tools for model extraction attacks.[Results/Conclusion] The experimental results show that model extraction based on the target distribution is feasible, and the resulting replacement model even outperforms the target model in some categories. Moreover, the effectiveness of the attack is enhanced as the distribution used approaches the target distribution.更多还原