Cover Reproducible Steganography via Deep Generative Models

Whereas cryptography easily arouses attacks by means of encrypting a secret message into suspicious form, steganography is advantageous for its resilience to concealing the in an innocent-looking cover signal. Minimal distortion steganography, one mainstream frameworks, embeds messages while minimizing caused modification on elements. Due unavailability original signal receiver, embedding realized finding coset leader syndrome function steganographic codes migrated from channel coding, which complex and has limited performance. Fortunately, deep generative models robust semantic generated data make it possible receiver perfectly reproduce stego With this advantage, we propose cover-reproducible where source e.g., arithmetic serves as code. Specifically, decoding process coding used encoding regarded extraction. Taking text-to-speech text-to-image synthesis tasks two examples, illustrate feasibility steganography. Steganalysis experiments theoretical analysis are conducted demonstrate that proposed methods outperform existing most cases.