Robust zero-watermarking based on dual branch neural network for ownership authentication, auxiliary information delivery and tamper detection

Abstract

This paper presents a robust multitask zero-watermarking scheme for ownership authentication, auxiliary information embedding, and tamper detection using a dual-branch neural network. The proposed method generates three zero-watermarking codes stored in a three-layer structure, where each layer corresponds to a different type of watermark: a binary logo for ownership authentication, a QR code for auxiliary data, and a halftone version of the original image for tamper detection. The first and third zero-watermarking codes are generated by a logical linking between the binary logo and halftone version, respectively, with a set of neural network weights. The second zero-watermarking code is created by linking the QR code with features extracted from the dual-branch neural network. This approach ensures that the original image remains undistorted and protected at the same time. Experimental results demonstrate the robustness of the proposed method against geometric distortions, common signal processing attacks, and combined attacks, achieving bit error rates below 0.005 and normalized correlation values close to or equal to 1. Additionally, the method attained an average accuracy of 98.7 % or higher in tamper detection tasks across multiple datasets, demonstrating its versatility.