Development of Multiple Forgery Detection and Localization Techniques for Digital Video

Abstract

In recent years, video footages have played an important role in evidence shreds in various sectors, including forensic, surveillance, social networks, courtrooms, and news media. However, with the unlimited availability of advanced editing software, it has become quite easy for even inexperienced users to change/modify the video’s actual content, resulting in an alarming rise in the number of forged videos. As a result, trust in videos is dwindling. There is a nudge to find a reliable solution to video forgery detection. Researchers across the world are working hard to come up with new solutions to the problem. There are two types of techniques for detecting video manipulation: active and passive. The active technique embeds footprint data in the form of watermarks in the digital video, either during recording or later, with the assistance of a specific application, allowing later verification of the originality of its content. However, the difficulty with the active method for video authentication is that it is only helpful in a restricted number of cases requiring special hardware. Passive techniques were developed in response to these constraints. It examines the integrity and authenticity of a video in the absence of an embedded watermark and relies on internal features. However, many of the passive approaches proposed so far can only identify one type of forgery at a time. This doctoral research aims to advance the field of video forensics by developing passive techniques for detecting and localizing the multiple forgeries in the forged video. In this study, first, the prerequisite information for comprehending video forgery detection is impersonated. Following that, a detailed review of other contemporary passive forgery detection techniques for digital video is conducted. Based on this review, the limitations of the existing work are identified. This study proposes techniques for passive video forgery detection. Each of these techniques concentrates on specific features or characteristics to identify the video forgeries. The first technique proposed in this work is based on correlation consistency between LBP-coded frames. This technique can detect and localizes the presence of forgeries in the videos and tested on an MFVD-4 dataset. This technique provides an accuracy of 97.33% and is robust against the illumination change, GOP length, and background dependency. This study offered a second passive technique to investigate multiple forgeries in a digital video using entropy-based texture features such as Two-Dimensional Distribution Entropy (DistrEn2D) and Bi-dimensional Multiscale Entropy (MSE2D), as well as the Outlier detection approach. This technique detects and locates the presence of forgeries when tested on an MFVD-2 dataset and provides an accuracy of 97.49% and 96.66% with DistrEn2D and MSE2D features, respectively. Moreover, the proposed technique is independent of GOP length and video background as well as robust to compression. The third passive technique detects multiple forgeries in a digital video using a VGG-16 neural network and Kernel Principal Component Analysis (KPCA). This technique achieves 97.24% accuracy when tested on an MFVD-3 dataset. Furthermore, this technique is independent of GOP length and video background and also robust to operations like noise addition and brightness, contrast, saturation, and hue modifications. Finally, a passive technique is proposed based on Principal Component Analysis (PCT) and Neighborhood Binary Angular Pattern (NBAP), followed by the GoogleNet model to detect and localize the multiple forgeries in the video. This technique has been tested and validated on the MFVD-1 dataset, and it is capable of identifying the forgeries in the video with 97.07% accuracy. Furthermore, the proposed technique is independent of GOP length and video background as well as robust to noise attacks. Experimental results demonstrate the efficiency of the proposed techniques. Comparisons with existing passive forgery detection techniques show improved performance of the proposed techniques.