Rad-Mark: Reliable adversarial zero-watermarking

Abstract

Zero-watermarking is a lossless protection technique, and thus it is widely used in medical images, artworks and other carriers that require lossless protection. However, the current zero-watermarking suffers from the problem of high similarity between the feature images of different host images, which results in a high false positive rate. To address this challenge, we propose Rad-Mark, a deep learning-based zero-watermarking framework that leverages adversarial feature optimization to enhance the robustness and accuracy of watermark detection significantly for the first time. The adversarial samples are employed to significantly improve the framework’s security, which can achieve the NC value of false positives close to 0.5. Both image perturbation and Gaussian noise are incorporated into the training process. Specifically, our Rad-Mark involves a feature fusion design, a mapping network based on the fusion of locally filtered and global handcrafted features. We conduct an in-depth analysis of key parameters, including Gaussian noise, watermark dimensions, and weighting factors, exploring their impact on the performance of our Rad-Mark. Extensive experimental results demonstrate that Rad-Mark outperforms existing zero-watermarking methods in terms of both security and robustness.