RoWSFormer: A Robust Watermarking Framework with Swin Transformer for Enhanced Geometric Attack Resilience

TL;DR

RoWSFormer tackles robust image watermarking by replacing CNN backbones with Swin Transformer blocks to capture global and multi-scale features, crucial for resisting geometric distortions. It introduces two novel blocks: Locally-Channel Enhanced Swin Transformer Block (LCESTB) for local and channel-aware feature extraction, and Frequency-Enhanced Transformer Block (FETB) to incorporate frequency-domain cues, all within an END watermarking framework. A constrained loss complements the standard encoder and decoder losses to keep watermarked pixels within valid ranges, and extensive experiments show superior PSNR and extraction accuracy, especially under rotations, scaling, and affine distortions, across non-geometric and geometric attacks. The results suggest RoWSFormer offers practical, high-fidelity watermarking with strong resilience to real-world geometric distortions, advancing the applicability of transformer-based watermarking in secure copyright protection.

Abstract

In recent years, digital watermarking techniques based on deep learning have been widely studied. To achieve both imperceptibility and robustness of image watermarks, most current methods employ convolutional neural networks to build robust watermarking frameworks. However, despite the success of CNN-based watermarking models, they struggle to achieve robustness against geometric attacks due to the limitations of convolutional neural networks in capturing global and long-range relationships. To address this limitation, we propose a robust watermarking framework based on the Swin Transformer, named RoWSFormer. Specifically, we design the Locally-Channel Enhanced Swin Transformer Block as the core of both the encoder and decoder. This block utilizes the self-attention mechanism to capture global and long-range information, thereby significantly improving adaptation to geometric distortions. Additionally, we construct the Frequency-Enhanced Transformer Block to extract frequency domain information, which further strengthens the robustness of the watermarking framework. Experimental results demonstrate that our RoWSFormer surpasses existing state-of-the-art watermarking methods. For most non-geometric attacks, RoWSFormer improves the PSNR by 3 dB while maintaining the same extraction accuracy. In the case of geometric attacks (such as rotation, scaling, and affine transformations), RoWSFormer achieves over a 6 dB improvement in PSNR, with extraction accuracy exceeding 97\%.

Figures (6)

• Figure 1: The difference between END-based model and flow-based model.
• Figure 2: The framework of RoWSFormer. The encoder $E$ and decoder $D$ consist of two crucial components: the Locally-Channel Enhanced Swin Transformer Block (LCESTB) and the Frequency-Enhanced Transformer Block (FETB). The encoder $E$ takes the cover image $I_{co}$ and watermark message $M_{en}$ as input and produces the watermarked image $I_{em}$ as output. The decoder $D$ receives the noise image $I_{no}$ as input and outputs the extracted watermark message $M_{ex}$. Between the encoder and decoder is a noise layer $N$, which includes both non-geometric and geometric distortions.
• Figure 3: The illustration of the Locally-Channel Enhanced Swin Transformer Block.
• Figure 4: The illustration of the Frequency-Enhanced Transformer Block.
• Figure 5: The watermarked image and the corresponding image with non-geometric distortions. Top: the cover image $I_{co}$; Second: the encoded image $I_{em}$; Third: the noise image $I_{no}$; Bottom: the residual image $|$$I_{em}$$-$$I_{no}$$|$.