高容量與穩健資訊隱藏技術:圖片及影片隱寫術

Abstract

圖片隱寫術是一種將機密資訊轉為微小波動再嵌入圖片的技術，傳統作法只能將極為少量的資訊藏於一張圖片上，近期使用深度學習的方式能將隱藏量提升10到100倍，使彩色圖片中的一個像素能藏入24位元的資料，但這種方式對於圖片的變動非常敏感，特別是JPEG壓縮的損壞會使藏入的資訊無法還原。而在影片隱寫術方面，一般是把影片中每一幀當作獨立的圖來套用圖片隱寫術，但這並沒有善用到每幀的時序特性，在本研究中，我們提出了兩種深度學習模型分別用於圖片和影片隱寫術，我們設計了端對端抗JPEG具注意力機制的模型(JASN)，它可以把一張機密圖像藏入另一張同等大小的圖片，即使該圖被JPEG壓縮過，我們仍然可以還原機密圖片。至於影片部分，我們設計一種高效的模型(FVSN)，在執行速度、模型大小和影片品質上，它的表現都比現存最好的模型還要好。

Image steganography is a technique to hide secret information into an ordinary image by adding visually indistinguishable perturbations. Traditional methods can only embed a small amount of data into the carrier image. The latest deep learning-based methods have achieved 24 bits payloads per pixel which is 10 to 100 times larger than typical methods. However, existing methods are sensitive to image distortion, especially the JPEG lossy compression, which makes models fail to reconstruct the secret. In video steganography, people usually extract the frames from videos and consider them as the image steganography task, but underutilize the temporal features. In this study, we proposed two deep networks for image and video steganography individually. We designed an end-to-end JPEG-resistant attention model, called JASN, which can robustly embed a secret image to a same-size cover image, and successfully reveal the secret image after processed by JPEG algorithms. We also designed an efficient video model, called FVSN, which outperforms the state-of-the-art model on speed, model size, and video quality.