TopStega—Generative Image Steganography Using Latent Space Techniques
Supervisor Name
Radi Jarrar
Supervisor Email
rjarrar@birzeit.edu
University
Birzeit University
Research field
Computer Science
Bio
I am an assistant professor of Computer Science at the Faculty of Engineering and Technology, Birzeit University, Palestine. Currently, I am the director of the PhD program in Computer Science at Birzeit University. I obtained my B.Sc. in Computer Information Technology from the Arab American University in 2007 and the Ph.D. from Monash University in 2012. My research interests include machine learning, data science, and computer vision, with applications in health informatics and computer security.
Steganography-the practice of concealing information within digital media-is vital for secure communication. Various approaches have been proposed yet they struggle to balance imperceptibility with reliable recovery. In this work, we present a novel generative image steganography framework that embeds binary messages directly into the latent noise tensors of text-to-image diffusion models. This aims at ensuring that hidden data coexists seamlessly with generative sampling. Unlike traditional hiding-in-pixels methods, our technique is undetectable by current steganalysis tools. Our pipeline dynamically adapts the spatial dimensions of the latent tensor to the message length and projects bits via a randomized keystream. During synthesis, we employ classifier-free guidance and a tailored denoising diffusion implicit models (DDIM) inversion to recover concealed information. We rigorously evaluate our method across numerous random seeds, demonstrating accurate extraction of message from generated images. This approach aims at advancing steganographic imperceptibility and robustness in generative AI and can be adapted in data transmission, provenance watermarking, and secure content sharing.
