Deep Robust Reversible Watermarking

TL;DR

This work proposes DRRW, a deep learning-based robust reversible watermarking framework that eliminates quantization-induced irreversibility by introducing an Integer Invertible Watermark Network (iIWN) built on Integer Coupling Layers. It employs an encoder–noise layer–decoder training scheme with a differentiable noise pool to learn robustness to distortions, and a two-stage inference pipeline that maps to an overflow stego image and a latent variable, whose auxiliary data are compressed via arithmetic coding and embedded with RDH for lossless recovery. Key contributions include an overflow penalty loss that reduces the auxiliary bitstream and improves image quality, and an adaptive loss-weight strategy that automatically tunes watermark robustness during training. Empirical results show DRRW achieves higher visual quality and robustness than state-of-the-art RRW methods, while delivering up to 55.14× embedding speedups and substantial reductions in auxiliary data, enabling scalable and practical deployment on large datasets such as PASCAL VOC 2012.

Abstract

Robust Reversible Watermarking (RRW) enables perfect recovery of cover images and watermarks in lossless channels while ensuring robust watermark extraction in lossy channels. Existing RRW methods, mostly non-deep learning-based, face complex designs, high computational costs, and poor robustness, limiting their practical use. This paper proposes Deep Robust Reversible Watermarking (DRRW), a deep learning-based RRW scheme. DRRW uses an Integer Invertible Watermark Network (iIWN) to map integer data distributions invertibly, addressing conventional RRW limitations. Unlike traditional RRW, which needs distortion-specific designs, DRRW employs an encoder-noise layer-decoder framework for adaptive robustness via end-to-end training. In inference, cover image and watermark map to an overflowed stego image and latent variables, compressed by arithmetic coding into a bitstream embedded via reversible data hiding for lossless recovery. We introduce an overflow penalty loss to reduce pixel overflow, shortening the auxiliary bitstream while enhancing robustness and stego image quality. An adaptive weight adjustment strategy avoids manual watermark loss weighting, improving training stability and performance. Experiments show DRRW outperforms state-of-the-art RRW methods, boosting robustness and cutting embedding, extraction, and recovery complexities by 55.14\(\times\), 5.95\(\times\), and 3.57\(\times\), respectively. The auxiliary bitstream shrinks by 43.86\(\times\), with reversible embedding succeeding on 16,762 PASCAL VOC 2012 images, advancing practical RRW. DRRW exceeds irreversible robust watermarking in robustness and quality while maintaining reversibility.

Figures (22)

• Figure 1: Comparison between traditional RRW and the proposed DRRW. (a) The traditional method embeds a robust watermark using conventional robust image watermarking methods; (b) The proposed DRRW embeds a robust watermark using an integer invertible watermark network.
• Figure 2: Comparison of revealed secret (without stego quantization) and Q-revealed secret (with stego quantization) in real-valued flow-based HiNet jing2021hinet.
• Figure 3: PSNR distributions of revealed secret (without stego quantization) and Q-revealed secret (with stego quantization) in real-valued flow-based HiNet jing2021hinet.
• Figure 4: Computation diagram of the forward and inverse mappings in the $i$th integer coupling layer of the integer invertible watermark network.
• Figure 5: The up/down-sampling network architecture of the DRRW framework. The up-sampling and down-sampling networks exhibit a symmetric structure, where the red arrows represent the down-sampling network and the blue arrows indicate the up-sampling network. Both networks have an initial convolutional layer, multiple sampling blocks, and a final convolutional layer.