Visual Decoding of Hidden Watermark in Trained Deep Neural Network
Shigeyuki Sakazawa, Emi Myodo, Kazuyuki Tasaka, Hiromasa Yanagihara · 2019 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR) · 2019
Deep neural network (DNN) is a kind of intellectual property considering its usefulness and cost to develop. This paper proposes watermarking to a trained DNN models to protect its copyright. The proposed method has a remarkable feature for watermark detection process, which can decode the embedded pattern cumulatively and visually. In the experiment, we can embed a specific visual pattern using 5,000 or 60,000 images on the pretrained image classification DNN model. Then the embedded pattern is decoded using 20 images out of the 5,000 or 60,000 images, while a performance degradation to the original image classification task is small. At the conference site, real time and animated visual decoding demonstration is performed.