A semi-fragile watermarking scheme based on neural network

A novel neural network based image watermarking algorithm in wavelet domain is described. We firstly select some wavelet coefficients in zerotrees from subimages of different resolutions according to different 2×2 pixel blocks, and then establish the relational model among these coefficients by using the neural network. Finally a bit of the watermark is embedded by adjusting the polarity between a high frequency coefficient and the output value of the model. In the proposed method, the filter banks are regarded as the key to take overall control of the embedding process, and so the algorithm is public while keeping high security. The experimental results show that the watermark can not only discriminate between malicious and incidental tamper but also exactly locate the malicious modifications.