A robust zero-watermarking technique for securing medical images using modified ResNet50 and Discrete Cosine transform

Abstract

With the improvement of mobile Internet technology, it has become important to secure medical image information. Conventional techniques for watermarking images face several challenges, particularly concerning the degree of security provided against attacks, which compromise the confidentiality and usability of medical images. To solve these issues, this article presents a novel zero-watermarking algorithm-based approach based on medical-image security. The proposed solution uses Discrete Cosine Transform (DCT) and a modified ResNet50 for feature extraction. Improved ResNet50’s last layer softmax and classification layer are completely replaced with a Fully Connected (FC) layer and a regression layer. The watermark content is embedded in the extracted features after first encrypting them in Arnold’s mapping and Logistic mapping to enhance security without altering the watermark data. The approach was tested on a publicly available Kaggle X-ray image dataset with five classes (Normal, Tuberculosis, COVID-19, Viral Pneumonia, and Bacterial Pneumonia), demonstrating a strong performance against Gaussian noise (0.01 to 0.09) and JPEG compression (5% to 25%). The algorithm achieved an average PSNR greater than 20 and an NC of 0.86, with robust results under geometric attacks. These findings highlight the effectiveness of the method in preserving the image quality and watermark resilience under various conditions.