Developing Forensic Analysis of Video Using Artificial Applications

Abstract

This thesis titled Developing Forensic Analysis of Video Using Artificial Application presents a significant advancement in video forensics by introducing a novel, object-based approach to forged frame detection using a specialized Convolutional Neural Network (CNN), referred to as the temporal-CNN. The proposed method focuses on identifying forged frames by analysing motion residuals, particularly targeting artifacts such as large, asymmetrical blobs indicative of manipulation. Unlike traditional methods, the temporal-CNN operates independently of the Group of Pictures (GOP) structure, enhancing its adaptability across various video codecs including MPEG-4 and H.264. Key contributions include the identification of an optimal temporal window size (19 frames) that yields the highest forged frame accuracy (FFAC) of 97.36% and a low probability of error (Pe = 0.0519). Comparative evaluations demonstrate the superior performance of the temporal CNN over steganalysis-based methods, including that of Chen et al., across single-class and multi-class classification tasks. The model effectively distinguishes between forged, double compressed, and pristine frames, achieving a double-compressed frame accuracy (DCFAC) of 98.94%. Comprehensive experiments on the SYSU-OBJFORG VFVL dataset comprising HD object based fake videos of varied frame sizes and motion complexity highlight the model s robustness. The temporal-CNN achieves an average F1-score of 0.8644 and significantly outperforms traditional feature-based approaches such as SPAM, CF*, and CCPEV. Further insights are gained through activation map analysis, revealing that the network learns to identify distinguishing tampering patterns. The system also demonstrates resilience against post-processing attacks such as frame rate alterations. Collectively, these findings underscore the temporal-CNN s effectiveness in detecting sophisticated video forgeries, representing a crucial step toward reinforcing the integrity and credibility of digital video evidence in forensic co