StegoZip: A Plug-and-Play Framework for Increasing Steganographic Payload Capacity with Large Language Model

ACL ARR 2025 February Submission4272 Authors

Abstract: Generative steganography is a current research hotspot, yet its secret message payload capacity is often limited by low entropy during generation. The low capacity necessitates long stego texts or numerous transmissions, increasing the risk of detection by third parties. Prior studies have primarily enhanced payload capacity by making more effective use of available entropy while largely overlooking the equally critical step of secret message preprocessing. In this paper, we propose StegoZip, the first plug-and-play framework that employs large language model-driven dynamic semantic redundancy pruning combined with index compression coding to optimize secret message preprocessing and further increase payload capacity. In combination with advanced steganography, the experimental results demonstrate that StegoZip can increase the payload capacity by 2–3× while reducing the time per unit message by approximately 50\%. Furthermore, StegoZip operates independently of the steganography embedding process, ensuring that it does not impact the security of the original method.

Paper Type: Long

Research Area: Information Extraction

Research Area Keywords: Information Extraction,Generation,Efficient/Low-Resource Methods for NLP

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings

Languages Studied: English,Binary

Submission Number: 4272