Real-time watermark reconstruction for the identification of source information based on deep neural network

A novel deep neural network-based image watermarking method is presented to identify the source of digital data that is shared/forwarded on the internet using various messenger apps. The app that is used to share/communicate the image at the very first time is also identified in the proposed method. The ten-digit mobile number of the source (user) and identification data of particular messenger app (i.e. WhatsApp, Snapchat, Kik, Facebook messenger, etc.) is combined to get the text watermark signal. The part of the watermark signal representing specific mobile-based messenger application is obtained by randomizing the Walsh orthogonal codes using secret keys. To embed the watermark, the host image (shared/forwarded) is divided into blocks of equal size and then, slantlet transform is applied on each block. To get high reliability, three copies of the source information (user and app) are embedded during watermark embedding. Watermark extraction is performed using trained multilayer deep neural network. Furthermore, an optimal block selection logic is used to get improved results for real-time applications. The method is examined against various signal-processing attacks and high robustness with significant imperceptibility is attained. The method is also found to be fast enough for real-time applications. The prime objective of identifying the first user (source) and the shared/forwarded status (app detection) is successfully accomplished.