Dataset authorization control: protect the intellectual property of dataset via reversible feature space adversarial examples

Training high performance Deep Neural Networks (DNNs) models require large-scale and high-quality datasets. The expensive cost of collecting annotating datasets make the valuable can be considered as Intellectual Property (IP) dataset owner. To date, almost all copyright protection schemes for deep learning focus on models, while is rarely studied. In this paper, we propose a novel method to actively protect from being used train DNN without authorization. Experimental results CIFAR-10 TinyImageNet demonstrate effectiveness proposed method. Compared with model trained clean dataset, effectively test accuracy unauthorized protected drop 86.21% 38.23% 74.00% 16.20% datasets, respectively.