Dual-Branch CNN-Transformer network for robust Zero-Watermarking of medical images

Abstract

In this paper, we propose a novel zero-watermarking approach that utilizes deep learning to protect medical images from malicious attacks, including unauthorized copying, cropping, and other forms of tampering, during transmission. In this paper, we created a two-part system that uses a simple Convolutional Neural Network (CNN) and a transformer to effectively understand both small details and the overall context of medical images, called a dual-branch CNN-Transformer network. The CNN branch extracts the local details while the transformer branch captures the global contextual information. The Maximum Voting Adaptive Fusion Module (MVAM) integrates these features to generate robust medical images representations. Logistics map encryption is used to ensure the integrity of the watermark without altering the original image. Experimental results show that the proposed method is robust to various attacks (including noise, compression, filtering, rotation, scaling, translation, and cropping) and outperforms existing techniques. The ability to extract up to 4096 dimensional features greatly improves the characterization of medical images and helps to improve disease diagnosis.