【论文阅读】Prediction Poisoning: Towards Defenses Against DNN Model Stealing Attacks（2020）

摘要

High-performance（高性能） Deep Neural Networks (DNNs 深度神经网络) are increasingly（越来越多） deployed（部署） in many real-world applications e.g., cloud prediction APIs. Recent advances in model functionality stealing attacks（模型功能窃取攻击） via black-box access（黑盒访问） (i.e., inputs in, predictions out 输入，预测输出) threaten（威胁） the business model（业务模型） of such applications, which require a lot of time, money, and effort to develop. Existing defenses（现有的防御） take a passive role（被动的作用） against stealing attacks, such as by truncating predicted information（截断预测信息）. We find such passive defenses（被动防御） ineffective（无效） against DNN stealing attacks. In this paper, we propose the first defense which actively pertu