Sstegan: Self-Learning Steganography Based On Generative Adversarial Networks

Steganography is designed to conceal a secret message within public media. Traditional steganography needs a lot of expert knowledge and complex artificial rules. To solve this problem, we propose a novel self-learning steganographic algorithm based on the generative adversarial network, which we called SSteGAN. This method learns the steganographic algorithm in an unsupervised manner without expert knowledge and directly generates the stego image from the secret message without the cover image. We define a game with four parts: Alice, Bob, Dev and Eve. Alice and Bob attempt to communicate securely. Eve eavesdrops on their conversation and wants to distinguish whether the secret message is embedded in the image. Dev attempts to determine real images from generated images. Experiment results demonstrate that Alice can produce vivid stego images and Bob can successfully decode the secret message with 98.8% accuracy.