Work

A Near-Imperceptible Disambiguating Approach via Verification for Generative Linguistic Steganography