摘要
Deep machine learning models(深度机器学习模型) are increasingly(越来越多) deployed(部署) in the wild(在野外) for providing services to users(给用户提供服务). Adversaries may steal the knowledge of these valuable models(窃取有价值模型的知识) by training substitute models(训练替代模型) according to the inference results of thetargeted deployed models(根据目标部署模型的推理结果). Recent data-free model stealing methods(无数据模型窃取方法) are shown effective(有效) to extract the knowledge of the target model(提取目标模型的知识) without using real query examples(不适用真实查询样本), but they assume(假设) rich inference information(丰富的推理信息), e.g., class probabilities and logits(类概率和逻辑). However, they are all based on(基于) competing generator-substitute networks(互相竞争的生成替代网络) and hence(因此) encounte