A Robustness Improved Watermarking Scheme based on Invertible Neural Network

Image watermarking technology aims to embed secret information imperceptibly into natural images for authentication or copyright protection. Invisibility and robustness are two primary concerns for watermarking algorithms. Previous methods leveraging invertible neural networks (INNs) have achieved high security in steganography. However, they lack resilience to image modifications, leading to compromised decoding quality. Conversely, encoder-decoder frameworks excel in robustness but often compromise invisibility. In this paper, a novel end-to-end robust watermarking framework is proposed that seamlessly integrates INNs with an encoder-decoder paradigm. Additionally, an efficient channel attention (ECA) block is incorporated in the message extraction stage, enhancing watermark extraction and consequently boosting decoding robustness. Furthermore, a message feature loss is introduced to regulate the INN, thereby accelerating network convergence and enhancing performance. Extensive experiments demonstrate that our proposed scheme outperforms existing algorithms in both robustness and invisibility. When subjected to a combination of 13 noise attacks, including Gaussian noise, random cropping, and JPEG compression, our model achieves a bit error rate below 1% with an embedding information length of 64 bits. Notably, the watermarked images maintain high visual quality, with PSNR and SSIM values reaching 36dB and 0.95 respectively.