Formal Security Analysis of CARTS Steganography in Autoregressive Models
September 11, 2026
This work provides the first formal security analysis of Contextual Autoregressive Rank Transcoding Steganography (CARTS), which encodes payloads into stegotext of identical token length. The paper characterizes security via context search, key collisions, and message equivocation within the rank-coordinate representation.
HOW THIS AFFECTS YOU
●
researcherYou can use these formal definitions to study the theoretical limits of steganographic capacity in LLMs.
●
policyThis research highlights how existing autoregressive models can be used for undetectable data encoding.