Robust image watermarking for diverse channels with template-forming neural network

Applied Soft Computing Journal. 2026. Vol. 186. Article 114125.

Dzhanashia K., Evsyutin O.

Watermarking is a key tool in combating unauthorized content distribution, but its effectiveness is often challenged by the wide range of communication channels that can degrade or remove the watermark. We propose a block neural network-based watermarking scheme for digital images that is robust against diverse transmission channels, including compression, pre-processing, and digital-to-analog conversions such as screen photographing and print-scan processes. Instead of using a more conventional encoder-decoder, a neural network with two inputs and inter-layer connectivity is used for forming block templates that are later superimposed into a cover image, while a classification neural network is used for extraction. The method achieves reliable extraction, with an average bit error rate below 20 % even in challenging conditions. It also preserves visual quality (PSNR > 30 dB) and supports a payload of 2 bits per 32 × 32 block, enabling watermark lengths sufficient for unique identification within a single image frame.

Keywords: data hiding, convolutional neural network, neural network, image watermarking, robust watermarking

Publication based on the results of: ­­­Efficient methods for embedding data into digital images using metaheuristic optimization (2024)