Robust image watermarking algorithm based on generative adversarial networks and dynamic attention

Image watermarking is a technique for protecting image copyright by embedding hidden information. In recent years, deep watermarking methods based on Invertible Neural Networks have gained attention due to their high capacity and bidirectional reconstruction performance. However, existing methods often compromise the invertibility of the model in the pursuit of robustness, making it easier for steganalysis networks to detect the presence of hidden information. To address this, we propose an image watermarking algorithm based on Generative Adversarial Networks (GANs) and dynamic attention. The algorithm constructs a GAN framework with a steganalysis network and an image embedding process, leveraging adversarial training to improve concealment. An image enhancement module with a dynamic attention mechanism is designed to adaptively focus on distorted regions and process different distortions in a categorized and phased manner, thereby enhancing robustness. Furthermore, a differential feature extraction module captures feature discrepancies between watermarked and attacked images, integrating these features into the extraction process to compensate for information loss and improve watermark quality. The core innovation of this work lies in the proposal of a collaborative three-layer watermarking framework that enables active concealment through adversarial training, adaptive resistance via dynamic attention mechanisms, and intelligent recovery using differential features, thereby collectively pushing the performance boundaries of existing solutions across multiple security metrics. Experiments on multiple datasets demonstrate that our approach significantly outperforms existing mainstream methods in extraction quality under various distortion attacks.