Robust Zero-Watermarking for Medical Images Based on Deep Learning Feature Extraction

Internet data transfer has increased in recent years, as well as the need to generate protection and authentication of multimedia data. The use of watermarks on digital images serves for copyright protection and ownership authentication. Zero-watermarking does not embed the watermark into the host image. The imperceptibility of the watermark is required for some tasks without modifying pixels on the image, such as medical images, because the medical image has a low distortion and may generate a wrong diagnosis. In this paper, we propose a zero-watermarking algorithm for medical images with patient authentication purposes and avoid image tampering based on a deep learning neural network model as a feature extractor using the Context Encoder. Thus, is extracted a feature map with the most representative image features. Therefore, an or-exclusive is applied to merge the watermark sequence and the extracted feature map. The watermark signal consists of a pseudorandom sequence and contains the patient's information. The bit error rate and the normalized cross-correlation demonstrate the robustness of the zero-watermark technique against geometric and image processing attacks.