Vision Transformer-Based Watermark Generation for Authentication and Tamper Detection Using Schur Decomposition and Hybrid Transforms

Abstract

Multimedia manipulation has increased which demand for security across various applications. Majorly image-oriented security issues such as image authentication, proof of ownership, and copyright protection are highly increased. To authenticate and detect the tamped region and to recover the tampered area vision transformer-based hybrid watermarking model is proposed. We proposed a novel model to achieve image authentication, tamper detection, and localization followed by image recovery. In the proposed model, invariant attention-based watermark feature maps are generated using a Vision transformer. We have generated three different watermarks: first using the SVD eigenvalue generated as an authentication watermark, secondly to detect tampered region average 6MSB of each 2*2 block generated by performing the Schur decomposition method on the biometric image, and to locate and recover the image Vision feature maps are generated and average 6MSB of each block is embedded as the tamper detection watermark. Normally the generated watermark is embedded either using an embedding factor or using a suitable embedding location. In the proposed model, watermark embedding is performed by finding the optimal embedding region using high entropy block region. On the original image, curvelet transform is performed followed by invariant integer wavelet transform. The first authentication eigenvalue is embedded on the LH band singular value diagonal matrix obtained by the SVD model. On the LL band, 2*2 blocks Schur decomposition is performed to embed the 6MSB in the 2LSB bit of upper triangular coefficient values. At last, vision feature maps are embedded in the curvelet approximate coefficient high entropy region and inverse curvelet performed, producing a watermarked image.