An Image Zero Watermark Algorithm Based on DINOv2 and Multiple Cycle Transformation

To address the issue of insufficient robustness in image feature extraction by current zero-watermarking techniques, which negatively impacts performance. This paper proposes an image zero-watermarking algorithm based on DINOv2 and Multiple Cyclic Transformations(MCT). The method first uses a pre-trained DINOv2 model to extract global features from the original image, which have invariance and generalization and can resist various image transformations and attacks. Secondly, it uses MCT to extract robust statistics (mean, variance, autocorrelation coefficient, etc.) from the image’s global features and constructs statistical features according to the statistics, which have good stability and anti-interference ability and can enhance the security and robustness of zero-watermarking. Thirdly, it binarizes the statistical features by threshold 0, obtains the image binary features, and combines them with the watermark sequence of the image owner by XOR operation, obtaining the zero watermark. Finally, it generates a zero watermark from the noisy image for copyright verification, which is symmetrical to the process of constructing a zero watermark from the original image. The experimental results show that the proposed zero-watermarking method is robust against different image attacks (such as rotation, scaling, filtering, compression, noise addition, etc.), and its average Bit Error Rate(BER) is 0.17%