doc. Ing. Tomáš Pevný, Ph.D.

doc. Ing. Tomáš Pevný, Ph.D.

Ing. Mgr. Ladislava Smítková Janků, Ph.D.

Department of Applied Mathematics

With new state-of-the-art autoregressive models, steganography now offers a way to hide messages using Large Language Models, while advertising provable security of these algorithms. This thesis explores the detectability of such existing steganographic algorithms by training a machine learning models, specifically Random Forest and Gradient Boosted Decision Trees in order to detect text with hidden messages. The obtained results show that the security of the final scheme is influenced not only by the algorithm itself, but also by other factors, such as the decoding methods, temperature or omitted tokens indicating a need for a more comprehensive approach evaluating security of such schemes.