Fundamental Tradeoffs for Information Hiding (DETERMINE)

How can data be securely hidden in data? This question has long been of interest to computer scientists and has taken on new significance with the breakthrough of generative artificial intelligence (AI). Generated media data, such as the output of AI chat bots, can carry hidden digital watermarks that can be used to trace the origin or prove the authenticity of the information. For example, breakthroughs in research could make it possible to automatically recognize texts generated by large language models and marked with digital watermarks with high confidence, even if they have been slightly edited. In addition, generative AI can also be used to introduce completely new types of watermarks into conventional media data. A precise analysis of the new, machine-learned watermarks promises new insights into how data can be hidden even more securely than previously known. Depending on the choice of target criteria in the learning phase, different properties can be realised, such as undetectability, robustness, or high capacity. Undetectability is a prerequisite for secure steganography for highly confidential communication. Steganography achieves even higher security than encryption because the very existence of the confidential message remains hidden. Secure steganography is the basis for censorship-resistant communication systems. Robustness is a prerequisite for the practical use of digital watermarks, which should still be readable even after changes to the medium (e.g., lossy transmission, processing by filters). The capacity indicates how much data can be hidden relative to the size of the cover medium. There is a conflict between undetectability, robustness and capacity. However, the exact relationships and limits are still unknown and need to be researched both in general and specifically for each media type (text, image, audio, compressed video). A particular challenge is to make the human perception of the modified media data measurable for computers. A research team at University of Innsbrucks Department of Computer Science, led by Univ.- Prof. Dr. Rainer Böhme, is working with scientists from the Czech Technical University in Prague to investigate the fundamentals of secure information hiding in the context of generative AI. The bilateral project, funded by the FWF and the GACR, is scheduled to run for three years and will advance the state of research in the fields of data and information security, signal processing, and machine learning.