Comparative Performance Analysis Of Deep Learning-Based Image Steganography Using U-Net, V-Net, And U-Net++ Encoders

Abstract

Digital Imaging steganography is the act of hiding information in a cover picture in a way that can't be found or recovered. Three main types of methods are used in digital image steganography: neural network methods, spatial methods, and transform methods. The pixel values of an image are changed by spatial methods to embed information. On the other hand, the frequency of the image is changed by transform methods to embed information that is hidden. There are methods that use neural networks to hide things, and this is what the suggested method is all about. Through digital image steganography, this study looks into how deep convolutional neural networks (CNNs) can be used. With the increasing concerns about data infringement during transmission and storage, image steganography techniques have gained attention for hiding secret information within cover images. Traditional methods suffer from limitations such as low embedding capacity and poor reconstruction quality. To address these challenges, deep learning-based approaches have been proposed in the literature. Among them, the Convolutional Neural Network (CNN) based U-Net encoder has been extensively studied. However, its comparative performance with other CNN-based encoders like V-Net and U-Net++ remains unexplored in the context of image steganography. In this paper, we implement V-Net and U-Net++ encoders for image steganography and conduct a comprehensive performance assessment alongside U-Net architecture. These architectures are utilized to conceal a secret image within a cover image, and a unified and robust decoder is designed to extract the hidden information. Through experimental evaluations, we compare the embedding capacity, stego quality, and reconstruction quality of the three architectures. The U-Net architecture outperforms V-Net and U-Net++ in terms of embedding capacity and the quality of stego and reconstructed secret images. This research provides valuable insights into the effectiveness of different deep learning-based encoders for image steganography applications, aiding in the selection of appropriate architectures for securing digital images against unauthorized access.