Dual Watermarking with Deep Learning for Enhancing Security of Medical Images

Rajat Sood, Jyoti Rani, Ashima Anand

Abstract: The rapid development of deep neural networks has prompted significant research into watermarking techniques for medical imaging based on deep learning. Ensuring copyright protection by watermarking medical images, particularly chest X-ray images is imperative. However, numerous extant methods cannot effectively utilize image characteristics to achieve embedding and extraction. This study introduces a new dual watermarking system that is based on deep learning and utilizes the Redundant Discrete Wavelet Transform (RDWT) to improve both imperceptibility and robustness. The proposed architecture utilizes a normalizing flow technique within the encoder to seamlessly embed a watermark into a medical cover image, resulting in a securely encoded output. Furthermore, the encoded image incorporates a sophisticated security layer to enhance its protection. The system’s robustness is further enhanced by adding a noise layer between the encoder and decoder, designed to separate the secret image with minimal loss. Our approach obtains watermark robustness and imperceptibility that are significantly superior to existing methods, as evidenced by extensive evaluations. Pixel loss can often be below 0.1%. Our method separates secret images with indistinguishable visual quality from the original secret image, as proven by experimental studies conducted on a ChestXray dataset obtained from NIH. The superiority of our dual watermarking architecture over contemporary methodologies is further validated by comparative analysis.

External IDs:dblp:conf/epia/SoodRA25