CAWNet: A Channel Attention Watermarking Attack Network Based on CWABlock

In recent years, watermarking technology has been widely used as a common information hiding technique in the fields of copyright protection, authentication, and data privacy protection in digital media. However, the development of watermark attack techniques has lagged behind. Improving the efficiency of watermark attack techniques and effectively attacking watermarks has become an urgent problem to be solved. Therefore, this paper proposes a watermark attack network called CAWNet. Firstly, this paper designs a convolution-based watermark attack module (CWABlock), which introduces channel attention mechanism. By replacing fully connected layers with global average pooling layers, the parameter quantity of the network is reduced and the computational efficiency is improved, enabling effective attacks on watermark information. Secondly, in the training phase, we utilize a largescale real-world image dataset for training and employ data augmentation strategies to enhance the robustness of the network. Finally, we conduct ablation experiments on CWABlock, attention mechanism, and other modules, as well as comparative experiments on different watermark attack methods. The experimental results demonstrate significant improvements in the effectiveness of the proposed watermark attack approach.