Jialong Zhang, Zhongshu Gu, Jiyong Jang, Hui Wu, Marc Ph. Stoecklin, Heqing Huang, Ian Molloy. Protecting Intellectual Property of Deep Neural Networks with Watermarking. AsiaCCS, 2018.

Zhang et al. propose a watermarking approach to protect the intellectual property of deep neural network models. Here, the watermarking concept is generalized from multimedia; specifically, the purpose of a watermark is to uniquely identify a neural network model as the original owner’s property to avoid plagiarism. The problem is illustrated in Figure 1. As watermarks, the authors consider perturbed input images. During training, these perturbations are trained to produce very specific outputs, as illustrated in Figure 2. For example, random pixels are added, or text is added to images. After training, the model can be uniquely identified by these perturbed watermark images that are unrelated to the actual task.

Figure 1: Illustration of the problem setting for watermarking.

Figure 2: Example watermarks.

Also find this summary on ShortScience.org.
What is your opinion on this article? Let me know your thoughts on Twitter @davidstutz92 or LinkedIn in/davidstutz92.