Copyright protection of deep neural network models using digital watermarking: a comparative study

Abstract

Nowadays, deep learning achieves higher levels of accuracy than ever before. This evolution makes deep learning crucial for applications that care for safety, like self-driving cars and helps consumers to meet most of their expectations. Further, Deep Neural Networks (DNNs) are powerful approaches that employed to solve several issues. These issues include healthcare, advertising, marketing, computer vision, speech processing, natural language processing. The DNNs have marvelous progress in these different fields, but training such DNN models requires a lot of time, a vast amount of data and in most cases a lot of computational steps. Selling such pre-trained models is a profitable business model. But, sharing them without the owner permission is a serious threat. Unfortunately, once the models are sold, they can be easily copied and redistributed. This paper first presents a review of how digital watermarking technologies are really very helpful in the copyright protection of the DNNs. Then, a comparative study between the latest techniques is presented. Also, several optimizers are proposed to improve the accuracy against the fine-tuning attack. Finally, several experiments are performed with black-box settings using several optimizers and the results are compared with the SGD optimizer.