Feature-Attention-Mechanism-Based Attack for Deep Robust Watermarking

Currently, deep neural networks have made significant progress in the attack of watermark removal. However, some challenges still exist, such as how to improve attack effectiveness while reducing the impact on image quality. In this article, we propose a feature-attention-mechanism-based attack scheme for deep robust watermarking (FAADW). To extract the residual feature map related to watermark message distribution, a residual attention block based on the feature attention mechanism is first designed. Then, we utilize a framework based on a conditional generative adversarial network, which is guided to use original images as labels to generate attacked images without sacrificing visual quality. A loss function that integrates adversarial loss, perceptual loss, and structural similarity loss is exploited to further supervise the training process of our FAADW. Experimental results show that the proposed scheme achieves satisfactory performances in terms of attack effectiveness for robust watermarking and visual quality of attacked images simultaneously.