A Master of Science thesis in Computer Engineering by Afaf Eltayeb Mohamedelbagir Ahmed entitled, “Data Embedding and Extraction in Scrambled Video using Machine Learning”, submitted in December 2020. Thesis advisor is Dr. Tamer Shanableh. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).
Data embedding in videos and images has various important applications such as digital rights management (DRM), content authentication, copyright protection, error resiliency and concealment as well as law enforcement. With the high possibility of illegal access and unauthorized content manipulation in shared storage platforms such as cloud data centers and with the risk of encountering different types of attacks during network transmission, videos and other sensitive data are usually transmitted and stored in an encrypted form. Accordingly, the need for data hiding techniques that operate directly on the encrypted video domain has emerged. This work proposes a novel data hiding scheme in encrypted video streams where scrambling and data embedding are performed simultaneously at the encoder side by rotating the motion vectors of the cover video. Then a machine learning solution is proposed at the decoder side to classify the motion vectors to rotated/ unrotated, extract the hidden information bits and reconstruct the original cover video. A sequence-dependent approach is applied where the first part of the video is used for training and model generation. The proposed system is composed of two phases: firstly, the training phase where the model is trained to distinguish between the correctly reconstructed macroblocks and the macroblocks reconstructed using rotated motion vectors. Secondly, the testing phase in which the trained model is applied to identify which of the candidate macroblocks are the ones associated with the true motion vectors. Once the true motion vectors are identified, they are compared to the ones received in the bit stream and thus the embedded bits are extracted, and the video is reconstructed. Experiments are conducted on a number of well-known video sequences after compressing them once with the Moving Pictures Expert Group-2 video codec standard and then with the High-Efficiency Video Coding standard. A detailed analysis is provided based on the macroblock type, the number of motion vectors and the type of the encoding sequence. Lastly, the proposed solution is evaluated in terms of classification accuracy, embedding capacity and reconstruction quality.