A Neural Watermark Approach

In this paper we propose the coupling of a watermarking technique for images, called least significant bit, in the multiple classes random neural network. For that, we design a training process of the watermark pattern, an embedding process of the learned pattern in the original image, and a detecting process of this pattern in the carrier image. The removal of the watermark is not considered in this work, since the aim is to study the capability of detection of our neural approach of any manipulation over the carrier image. We define several attacks to compare the robustness of our approach with previous works (rotation, JPEG compression, scaling, noise, cropping, Horizontal Flip, Brightness and contrast correction), obtaining very good results with our approach. Additionally, we obtain very good performances in terms of the Peak Signal to Noise Ratio and Noise Generated criteria.