A Master of Science thesis in Computer Engineering by Seba Youssef entitled, “Detection of Double and Triple Compression in Videos for Digital Forensics 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).
Digital video forensics is the process of analysing, examining and comparing a video for use in legal matters and court cases. In digital video forensics, the main aim is to detect and identify video forgery and manipulation to ensure a video’s authenticity and reliability for use in court. This work focuses on passive forensics techniques, namely compression-based digital video forensics. When a video is edited by methods such as frame deletion, cropping, or duplication, the original encoded bitstream is first decoded, editing is applied and then the video is re-compressed before saving it. This means that by detecting re-compression in videos, we can interpret that the video has undergone some form of manipulation. The least number of recompressions a video can have is double compression, the first results from the device initially capturing the video which compresses it to store it in a suitable format and the second comes from the editing software or tool that re-compresses the video after it has been edited. Such editing can also be done multiple times leading to multiple compressions. Thus, finding out the compression history of a video becomes a very important mean for detecting any manipulation. Several techniques have been studied and investigated for the accurate classification of double and triple compression in videos based on machine learning and deep learning models with promising results being obtained. In this work, a number of experiments are conducted by using K-Nearest Neighbours (KNN), Random Forest (RF) or bi-directional Long Short-Term Memory (bi-LSTM) classifiers on a dataset of forged and unforged video sequences. In each of the experiments, performance is evaluated based on the classification accuracy and confusion matrix. Experiments are conducted on MPEG2 and HEVC coded videos using the same re-compression quantization parameter and the results of recompression detection are compared. Experiments are also conducted on HEVC coded videos with the same recompression bitrate and the results obtained are compared to existing solutions in literature. The experimental results revealed that both double compression and triple compression can be accurately detected using the proposed machine learning and deep learning solutions.