Seung-Min Mun

Wook-Hyung Kim, Jong-Uk Hou, Seung-Min Mun et al.

Since real-time contents can be captured and downloaded very easily, copyright infringement has become a serious problem. In order to reduce the loss caused by copyright infringement, copyright owners insert a watermark in the content to protect the copyright using illegal distribution route tracking and copyright authentication. However, whereas many existing watermarking techniques are robust to signal distortion such as compression, they are vulnerable to geometric distortion that causes synchronization errors. In particular, capturing real-time content in Internet browsers and smartphone applications is problematic because geometric distortion such as scaling and translation frequently occurs. In this paper, we propose a convolutional neural network-based template architecture that compensates for the disadvantages of existing watermarking techniques that are vulnerable to geometric distortion. The proposed template consists of a template generation network, a template extraction network, and a template matching network. The template generation network generates a template in the form of noise and the template is inserted into certain pre-defined spatial locations of the image. The extraction network detects spatial locations where the template is inserted in the image. Finally, the template matching network estimates the parameters of the geometric distortion by comparing the shape of spatial locations where the template was inserted with the locations where the template was detected. It is possible to recover an image in its original geometrical form using the estimated parameters, and as a result, watermarks applied using existing watermarking techniques that are vulnerable to geometric distortion can be decoded normally.

Seung-Min Mun, Seung-Hun Nam, Han-Ul Jang et al.

This paper introduces a blind watermarking based on a convolutional neural network (CNN). We propose an iterative learning framework to secure robustness of watermarking. One loop of learning process consists of the following three stages: Watermark embedding, attack simulation, and weight update. We have learned a network that can detect a 1-bit message from a image sub-block. Experimental results show that this learned network is an extension of the frequency domain that is widely used in existing watermarking scheme. The proposed scheme achieved robustness against geometric and signal processing attacks with a learning time of one day.