Watermarking Using Neural Network and Hiding the Trained Network within the Cover Image

The major source of communication in present world is digital media. As it is quite easy to manipulate a digital media, it becomes essential to protect the digital information by legitimate means. Digital Watermarking has evolved as one of the latest technologies for digital media protection. Many techniques based on spatial and frequency domain have been developed in the recent past and are being used for effective watermarking. However, there is always a tradeoff between robustness and imperceptibility features of watermarking offered by these techniques. This paper offers a technique based on Backpropagation Neural Network to train a given cover image to produce a desired watermark image. At the end of the training, the entire trained neural network weights has been successfully hidden within the cover image itself. This makes it possible to supply only the cover image without any external weight files. By extraction techniques, the weights can be derived from the cover image and used to reconstruct the trained Neural Network again which in turn converts the cover image into desired watermark image. The technique for hiding the weights into the cover image has been designed in such a way that it does not produce visual deterioration of the original cover image. This method is extra secure as it leads to watermarking indirectly.