Towards High-Capacity Provably Secure Steganography via Cascade Sampling

Meiyang Lv, Haocheng Fu, Xiaowei Yi, Hongxian Huang, Yun Cao, Changjun Liu

Abstract: The autoregressive large language model makes it possible to achieve provably secure steganography by controlled and accurate sampling of explicit distributions. However, existing provably secure steganography schemes focus on the construction of secure methods and neglect the consideration of steganographic capacity. To address this limitation, this paper introduces CascSamp, a high-capacity provably secure steganography method. CascSamp not only achieves higher embedding capacity, but also maintains superior security, efficiency, and generation quality. Firstly, CascSamp embeds secret information through a recursive uniform grouping sampling approach for tokens. It further achieves fine-grained controllable uniform grouping by increasing the sampling depth. As a result, each token can carry information more efficiently, which improves the embedding capability. Secondly, CascSamp jointly samples multiple tokens. This increases the amount of information that each token can carry, thus realizing higher embedding capacity. Finally, during the sampling process, CascSamp maintains the consistency of the original conditional distribution. This ensures provable security. Experimental results show that CascSamp achieves the same security compared to the state-of-the-art methods. It improves the embedding rate and the embedding speed. At the same time, CascSamp has lower computational complexity with generating high-quality stego images stably.

External IDs:dblp:conf/icics/LvFYHCL25