摘要
Training deep neural networks(训练深度神经网络) requires(需要) significant computational resources(大量计算资源) and large datasets(大型数据集) that are often confidential(机密的) or expensive(昂贵的) to collect. As a result(因此), owners(所有者) tend to(倾向于) protect(保护) their models by allowing(允许) access(通过) only via an API. Many works demonstrated(展示) the possibility(可能性) of stealing(窃取) such protected models by repeatedly querying(反复查询) the API. However, to date(到目前为止), research has predominantly(主要) focused on(集中) stealing classification models(窃取分类模型), for which a very large number of queries(大量的查询) has been found necessary. In this paper, we study the possibility of stealing image-to-image models(窃取图像到图像模型的可能性). Surprisingly(令人惊讶的是), we find(发现) that many such models can be stolen with as little as a single, small-sized, query image using simple distillation(使用简单蒸馏). We study this phenomenon(现象) on a wide variety(各种各样) of model architectures(模型架构), datasets(数据集), and tasks(任务), including denoising(去噪), deblurring(去模糊), deraining(去训练), super-resolution(超分辨), and biological image-to-image translation(生物图像到图像的转换). Remarkably(值得注意的是), we find(发现) that the vulnerability to stealing attacks(窃取攻击) is shared by CNNs and by models with attention mechanisms(注意力机制), and that stealing is commonly possible even without knowing the architecture of the target model(即使不知道目标模型的体系架构).
论文链接
Stealing Image-to-Image Translation Models With a Single Query