Online demo

arXiv link

Code

We recently released our demo for our EMNLP paper “Neural Linguistic Steganography”, hiding secret messages in natural language via arithmetic coding and GPT-2. Arithmetic coding is a powerful entropy coding technique that is optimal for random sequences. Using arithmetic coding in reverse enables extremely efficient steganography, and when combined with modern language models like GPT-2 it allows for convincing cover text generations that encode information.