A new approach to video steganography models with 3D deep CNN autoencoders.

This paper deciphers the sail on a proposition of a video steganography model with the use of a 3D DeepCNN- grounded autoencoder which allows extracting spatiotemporal features from still frames. It tends to hide frames of one video in another video taking into consideration equivalency in terms of size. For the phase of training the model, we've employed a UCF101 dataset containing 101 classes of action recognition videos (13320 videos). The quantitative benefaction of this model was arranged using varied quantitative indexes (SSIM, APD, and PSNR), and the qualitative benefaction was evaluated against the existing approaches. [ABSTRACT FROM AUTHOR]