A ROBUST MULTI-VIEW ZERO-WATERMARKING SCHEME USING SELF-SUPERVISED VISION TRANSFORMERS AND KEYED RANDOM HYPERPLANE HASHING

Abstract

This paper proposes a multi-view zero-watermarking scheme based on deep semantic features and keyed random hyperplane hashing to enhance robustness, discrimination, and security. Global features are extracted from the classification token of the self-supervised Vision Transformer DINOv2 and binarized using a Rademacher random projection matrix. The projection matrix is generated from a secret key, image identity, view index, and a 64-bit perceptual hash, forming an image-adaptive keyed hashing mechanism. To improve robustness against translation and cropping, zero-watermarks are generated from five sub-views and fused by majority voting. For geometric distortions, a spatial alignment procedure is applied before feature extraction. Experiments on biomedical and natural image datasets demonstrate strong robustness under common attacks while maintaining good discrimination, confirming the effectiveness and security of the proposed scheme for practical image copyright protection.