Text steganography on RNN-Generated lyrics.

Area of Science:

• Natural Language Processing
• Steganography
• Artificial Intelligence

Background:

• Traditional steganography methods often lack naturalness and have limited embedding capacities.
• Generating creative text, such as song lyrics, for steganographic purposes presents unique challenges in maintaining coherence and naturalness.

Purpose of the Study:

• To develop a novel method for embedding secret information within Chinese pop music lyrics.
• To enhance the embedding capacity and naturalness of steganographic messages hidden in generated text.

Main Methods:

• Utilized a Recurrent Neural Network (RNN) Encoder-Decoder architecture, specifically employing a Long Short-Term Memory (LSTM) model.
• Generated Chinese pop lyrics character by character or word by word, conditioned on an initial line.
• Incorporated common lyric formats and rhyme schemes to ensure naturalness and meet specific embedding requirements.

Main Results:

• The proposed RNN Encoder-Decoder model successfully generated Chinese pop lyrics embedded with secret information.
• Experimental and theoretical analyses demonstrated higher embedding capacities compared to existing text-generation steganography methods.
• The generated lyrics were found to be more natural-looking and sounding than those produced by prior art.

Conclusions:

• The RNN Encoder-Decoder model offers an effective approach for steganography in Chinese pop lyrics.
• This method achieves a superior balance between embedding capacity and the naturalness of the hidden message.
• The findings suggest potential for more sophisticated and covert information hiding in creative text generation.