Dynamic Codebook and Multimodal LLM Improve Text Steganography Security
A new black-box text steganography method using a dynamic codebook and multimodal large language model (MLLM) enhances security and practicality. Existing white-box methods risk exposure by sharing a language model between sender (Alice) and receiver (Bob), while black-box methods require a fixed codebook and specific extraction prompts, reducing flexibility. The proposed approach constructs a dynamic codebook via shared session configuration and an MLLM, then applies encrypted steganographic mapping to embed secret messages during caption generation. The method is detailed in arXiv:2604.20269.
Key facts
- arXiv:2604.20269 introduces a black-box text steganography method.
- The method uses a dynamic codebook and multimodal large language model.
- White-box steganography risks exposure due to shared language model.
- Black-box methods lack flexibility with fixed codebooks and extraction prompts.
- Dynamic codebook is constructed via shared session configuration and MLLM.
- Encrypted steganographic mapping embeds secret messages during caption generation.
- The approach aims to improve security and practicality.
- The paper is a cross-type announcement on arXiv.
Entities
Institutions
- arXiv