MTS-Stega: Linguistic Steganography Based on Multi-Time-Step.

Area of Science:

• Natural Language Processing
• Cryptography
• Machine Learning

Background:

• Generative linguistic steganography embeds confidential messages within text generated by language models.
• Current methods, primarily fixed-length and variable-length coding, face challenges with text quality and concealment.
• Fixed-length coding offers simplicity and low computational overhead, making it suitable for resource-constrained environments.

Purpose of the Study:

• To address the limitations of existing generative text steganography methods, specifically the reduction in text quality and concealment.
• To propose a novel steganography method that improves the selection of words to carry secret information while maintaining the statistical distribution of the generated text.

Main Methods:

• The proposed method utilizes a multi-time-step approach for word selection in generative steganography.
• This technique integrates multiple time steps to ensure selected words are both informative for secret messages and statistically consistent with the training data.
• The GPT-2 language model was employed for text generation in experimental validation.

Main Results:

• The multi-time-step method effectively improves the quality and concealment of steganographic text compared to conventional single-time-step approaches.
• Theoretical analysis and experimental results demonstrate the efficacy of the proposed scheme.
• The method successfully balances the embedding of secret information with the naturalness of the generated text.

Conclusions:

• The developed multi-time-step generative linguistic steganography method offers a significant improvement over existing techniques.
• This approach enhances text quality and concealment, making it a promising solution for secure communication in various applications.
• The findings highlight the potential of multi-time-step selection for advancing the field of steganography.