Robust deep image-watermarking method by a modified Siamese network

The risk of copyrighted data theft has increased with the growth of digital communication and easy access to digital data. Using watermarking methods has always been an active research subject against unauthorized users to protect copyrighted data. Deep watermarking methods were introduced in recent years to face this challenge, however, the robustness of these methods against graphic attacks and generating high-quality marked-images are still one of the important challenges in this issue. Also, exposure of these methods to the risk of intellectual property infringement is another important challenge in this subject. To address the quality degradation challenge of marked-images, we propose a blind deep image-watermarking method inspired by the Siamese network in which the watermark-codes will be embedded in the network’s weights instead of cover-images. Also, a cover-atmosphere is defined for each cover-image which includes the cover-image and its various attacked versions. Favorable robustness is achieved by mapping each cover-atmosphere to its corresponding watermark-code space. Moreover, an independent subspace for non-watermark-images is considered to map into a null-code which makes the method robust against intellectual property infringement attack. The results obtained from the experiment show that the suggested approach outperforms existing methods and can withstand different types of graphical and surrogate model attacks.