Perceptual Authentication Hashing for Digital Images With Contrastive Unsupervised Learning
Guopeng Gao, Chuan Qin, Yaodong Fang, Yuanding Zhou
Abstract
In recent years, many perceptual image hashing schemes for content authentication have been proposed based on classical methods and deep learning. However, most existing schemes target specific and limited content-preserving manipulations and cannot provide satisfactory robustness to unknown manipulations. In this work, we propose a new perceptual authentication hashing model for digital images based on contrastive unsupervised learning. In detail, a contrastive augmentation structure is exploited, which can optimize the model through changing the types and strengths of sample augmentation. Also, an integrated loss function is designed by the weighted summing of two components, i.e., the contrastive loss and hash loss, which can help the model learn perceptual feature representation with an unlabeled dataset and effectively improve the robustness and discrimination. Experimental results show that the proposed scheme can achieve superior performance compared with some state-of-the-art schemes, especially robustness to unknown attacks.