Multi-image Feature Map-Based Watermarking Techniques Using Transformer

Abstract

Nowadays, protecting multimedia data is a significant challenge because of the advancement of technology and software. The embedding process heavily relies on watermarking to accomplish multimedia security in terms of content authentication, proof of ownership, and tamper detection. Our objective is to develop an invariant watermark that can survive different signal-processing attacks. We presented a unique hybrid technique (DWT-QR-SWT) and multi-image invariant features generated as a watermark using a Transformer encoder-decoder model. The encoded image features are subsampled using PCA in order to decrease the dimensionality of the watermark image. The first two images are used as watermark1 and the next two images as watermark2 to produce multi-watermark feature maps. To embed the watermark, a hybrid DWT-QR decomposition has been applied to the original image1. On the primary watermarked image, two Level Stationary Wavelet Transform (SWT) were applied to embed the secondary watermark2. At the extraction phase, the tampered image is recovered by passing the extracted watermark image as input to the transformer decoder. A multi-image watermark increases data embedding capabilities and also achieves two-level content authentication, tamper detection, localization, and recovery. With a PSNR of 59.05 dB, the testing result demonstrates great resilience and improved imperceptibility.