Neural Linguistic Steganography

Zachary M. Ziegler*, Yuntian Deng*, Alexander M. Rush

EMNLP-IJCNLP 2019 · * Equal contribution · First preprint: September 3, 2019

Hide a message in generated natural language using arithmetic coding and a pretrained language model. STEGASURAS is the interactive demonstration of our 2019 research.

LM context ⓘ Context/history for language model generation, shared between sender and receiver. The cover text will follow directly after. Secret message ⓘ Message text to be concealed.

ⓘ How message text is converted into bits, affects cover text length. Message encode algorithm:
ⓘ How bits are converted into cover text. Each algorithm has its own parameter controlling tradeoff between compression and quality. Steganography algorithm:
ⓘ Temperature modules logits, lower is better quality. LM temperature:
ⓘ Remaining lowest probability tokens are truncated, lower is better quality. Truncation length:
ⓘ Bins to split vocab into, lower is better quality. Num bins: