Rewriting-Stego: Generating Natural and Controllable Steganographic Text with Pre-trained Language Model

Fanxiao Li, Sixing Wu, Jiong Yu, Shuoxin Wang, BingBing Song, Renyang Liu, Haoseng Lai, Wei Zhou

Abstract: Data transmission security and privacy play a crucial role in the era of information technology. Although the widely-used data encryption technique can ensure security, it can be easily detected and blocked by the observation system because the encrypted data format is quite different from the normal data. This work focuses on linguistic steganography, hiding a secret text in another normal stego text to ensure security and decrease the risk of being detected simultaneously. Rather than following the existing edit-based or generation-based paradigm, we propose a novel rewriting-based Rewriting-Stego, which tries to hide a secret text in the stego text by rewriting the given cover text. This paradigm integrates the advantages of both the edit-based paradigm and the generation-based paradigm, bringing higher information capacity without losing naturalness and controllability. Extensive experimental results on three public datasets have demonstrated the effectiveness of our Rewriting-Stego in terms of multiple metrics.