Zero Watermarking Using Convolutional Additive Self‐Attention Vision Transformer and Discrete Wavelet Transform‐Variance‐Based Feature Descriptor for Medical Image Security in Mobile Healthcare Services

To achieve robust and efficient copyright authentication for medical image in mobile e‐health services, a zero‐watermarking algorithm based on the architecturally refined convolutional additive self‐attention vision transformer (CAS‐ViT) and the discrete wavelet transform (DWT) variance‐based feature descriptor (DVFD) is proposed. The architecturally refined CAS‐ViT serves to extract deep feature maps from medical images, enhancing robustness while maintaining computational efficiency. The DVFD, which integrates DWT, variance pooling, and maximum election statistics, is employed with average hash (aHash) to generate a binary feature vector from a sequence composed of these extracted feature maps, further enhancing robustness. In the DVFD, DWT first decomposes each feature map into one low‐frequency approximation sub‐band and three high‐frequency detailed sub‐bands, variance pooling is then employed to calculate the variance of each high‐frequency sub‐band. Afterward, three DWT high‐frequency variance sequences are constructed, and maximum election statistics are employed with aHash to encode those sequences into a binary feature vector. To enhance watermark security, a control‐parameter‐range‐optimized Logistic mapping chaotic system is employed for watermark permutation. Experiment results demonstrate that this algorithm achieves excellent comprehensive performance of robustness and computational efficiency.