Image steganography using U-Net architecture supported by residual blocks and analysis of hidden data size

In modern communication systems, data security is of paramount importance. The primary goal is to ensure that sensitive information is transmitted to the intended recipient securely and unintelligibly to unauthorized individuals. Advancements in digital technology and the proliferation of devices have introduced new challenges in data security. In fields such as banking, healthcare, and personal privacy, the importance of data security has become increasingly critical. In this context, methods of data concealment like steganography have gained prominence for their ability to protect against malicious access. Steganography, by discreetly embedding crucial information within digital media, ensures that the data is only known to the sender and the receiver, making it a frequently employed method in the field of information security. This thesis is primarily focused on employing the U-Net architecture, which is supported by residual blocks, for the efficient concealment of colored message images of 256x256 dimensions within cover images of identical size. The classical U-Net architecture, traditionally used for image segmentation in the literature, has been adapted in this study for data hiding and extraction. During the testing phase of the model, two distinct analyses were conducted. Differing from existing studies, the first analysis investigated the impact of colored message images of various sizes (32x32, 64x64, 128x128, and 256x256) on the cover image using the Linnaeus 5 dataset. The second analysis aimed to measure the generalization capability of the model on images with different characteristics, employing additional datasets such as ImageNet and Labeled Faces in the Wild (LFW), and the results were compared with other studies in the literature. Comprehensive analyses have shown promising results in terms of Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) compared to current deep learning algorithms in the literature, to the best of our knowledge. The results demonstrate both a high capacity for data concealment and a high level of imperceptibility. Additionally, as part of the thesis work, cover images categorized based on their complexity levels and evaluates the measurement results obtained by embedding the same secret images into these two different categories. This provides a statistical assessment for selecting the optimum cover image based on complexity level.