The Influence of Neural Network Image Compression Methods on Digital Watermarks

The relevance of studying the effect of image compression using machine learning on digital watermarks is based on the high efficiency of neural network compression methods and the need to determine their effect on embedded information. In this paper, we investigate the measure of the influence of machine learning-based compression methods on embedded digital watermarks. Two approaches to embedding watermarks have been implemented: embedding into a spatial area and a transformation area. The systems underlying modern embedding systems were chosen as information embedding systems. The machine learning-based compression method taken into consideration justified the effectiveness claimed by the authors. The conducted computational experiments indicate the lack of resistance to compression using machine learning of information embedding systems based on modification of the spatial domain. Embedding systems that change the domain of the discrete wavelet transform (DWP) also do not have stability at compression rates exceeding the established “threshold” compression level, and demonstrate stability under other circumstances. Machine learning-based compression resistance is demonstrated by SSW and DWT-SVD systems.