CoPaD-Mark: A Coded Parallelizable Deep Learning-based Scheme for Robust Image Watermarking

In this work, we propose a coded, parallelizable, and deep learning-based image watermarking scheme (CoPaD-Mark). The embedding and extraction layers both undergo iterative training to achieve robust image watermarking. The embedding layer employs a parallel structure with convolutional neural networks inspired by the Inception Net, while the extraction layer uses deformable convolutions. These convolutions enable dynamic feature extraction by adjusting convolutional filters based on image content. The information bits embedded in the image are derived from the parity bits of an error-correcting code. This code’s message is constructed by combining a chaotic binary sequence with the original watermark bits. The inclusion of chaotic sequences adds randomness and complexity to the encoding process, enhancing robustness against attempts to alter the watermark. The proposed method produces watermarked images with peak signal-to-noise ratios (PSNR) above 36 dB. CoPaD-Mark is evaluated under several common attacks, including noise addition, filtering, cropping, compression, and resizing, and demonstrates robust performance compared to state-of-the-art methods.