Steganography method
AssigneeDefense Agency for Technology and Quality
InventorsJin Woo Ahn, Sang Woo Noh, Tae Hwan Kim
Life of this patent
Estimating the term…
Abstract
A steganography method and an apparatus embedding a hidden message in a video stream of an original cover video to produce a stego video using a generative neural network. The original cover video is a video stream as opposed using an image. A discriminant neural network is used to correct a generated video compared to an original cover video. The generated neural network is trained when the generated video is determined to be real by the discriminant neural network. A second stego video is acquired by adversarial training through a mutually adversarial relationship between a generative neural network and an discriminant neural network. The embedding technique is one of pre-embedding, intra-embedding, or post-embedding.
Claims9
A steganography method comprising:
- a first generating neural network application step of embedding an original hidden message in a video stream of an original cover video to generate a first stego video by applying a first generating neural network thereto;
- a first discriminant correction step of comparing the first stego video generated in the first generating neural network application step with the original cover video two or more times by using a first discriminant neural network, and correcting it;
- a first learning step of training the first generating neural network so that the first stego video generated in the first generating neural network application step is determined to be a real video in the first discriminant correction step;
- a second learning step of training the first discriminant neural network so that when the original cover video is input in the first discriminant correction step, it is determined to be real, and when the video derived from the first generating neural network is input therein, it is determined to be fake;
- and
- a second stego video acquisition step of acquiring a second stego video by using a first generative adversarial neural network induced (adversarial training) through a mutual adversarial relationship between the first generating neural network and the first discriminant neural network;
- wherein the original cover video is a video stream comprising any one or more of images and sounds;
- wherein the embedding method is one of pre-embedding, intra-embedding, and a post-embedding.
Dependencies are inferred from the claim text.