K-008
A Parameters Sharing Watermarking Technique for Convolutional Neural Network Model

◎Han He・Seok Kang・Yuji Sakamoto（Hokkaido Univ.）

Convolutional Neural Network (CNN) has made significant progress in various fields. Due to the high training cost, a high-performance CNN model can be considered as an Intellectual Property (IP). However, it may be illegally copied or redistributed. In this paper, we proposed an improved watermarking technique for CNN models. By sharing partial parameters of the host network to a multilayer perceptron, the embedding and training are done respectively. The experimental results show satisfactory performance in terms of security, robustness and embedding capacity compared to the conventional CNN watermarking scheme.