A neural network approach to digital data hiding based on the perceptual masking model of the human vision system

The purpose of this paper is to present a novel approach for digital watermarking and steganography technique that uses neural networks. The performance of the proposed solution in terms of its capacity, transparency, and robustness is investigated.

The proposed technique trains a neural network to learn the perceptual masking model of the human vision system. Once trained, the neural network identifies pixels whose most significant alteration will be least perceptible to the human eye. The image is then altered based on the network recommendation to include the watermark or the covert data.

Experimental results demonstrate that the proposed technique offers excellent transparency and good capacity. In addition, since neural networks store their learned knowledge in a distributed fashion, steganalysis of the image without access to the network is very difficult, if not impossible. Results demonstrate good performance of the proposed solution in terms of its capacity, transparency, and robustness.

Use of neural networks in extracting and representing perceptual masking model of human vision system is interesting. Value added by the proposed approach is in its use of artificial neural networks to model the perceptual masking model of human vision system for injecting imperceptible data into most perceptually significant pits of an image. The proposed approach may be used in combination with most current and popular methods with little impact on perceptual quality of the resulting image.