Digital Media Forgery Detection

In recent years, digital media have been used as indisputable evidence of a crime, and it is therefore important to ensure the reliability of these digital media. Unfortunately, digital media can be forged by using the advent of powerful and easy-to-use media editing tools. A forged digital medium is often eye-deceiving and appears in a way that is realistic, hence believable. A fundamental challenge is to determine whether a digital medium is authentic or not. This task is particularly challenging due to the lack of ground truth bases that can be used to verify the originality and integrity of digital media content. In this thesis, we investigate digital media forgery, particularly image and video forgery, and propose novel approaches to detect and localize digital media forgery. The first part of the thesis is devoted to image forgery detection. We propose a novel approach that uses dictionary learning and sparse coding to detect digital image forgery. We also propose a new matching criterion that is performed using dictionary atoms instead of traditional matching criteria. We conduct our experiments using two popular data sets to determine how effectively and efficiently our approach detects digital image forgery compared to previous approaches. The experimental results show that our approach outperforms state-of-the-art approaches and leads to robust results against compression and rotation attacks. Furthermore, our approach detects forgery significantly faster than these approaches since it uses a sparse representation that dramatically reduces dimensions of feature vectors. The second part of the thesis is devoted to object removal video forgery detection. We propose a novel approach based on sequential and patch analyses to detect object removal forgery and to localize forged regions in videos. Sequential analysis is performed by modeling video sequences as stochastic processes, where changes in the parameters of these processes are used to detect a video forgery. Patch analysis is performed by modeling video sequences as a mixture model of normal and anomalous patches, with the aim to separate these patches by identifying the distribution of each patch. We localize forged regions by visualizing the movement of removed objects using anomalous patches. We conduct our experiments at both pixel and video levels to determine the effectiveness and efficiency of our approach to detection of video forgery. The experimental results show that our approach achieves excellent detection performance with low-computational complexity and leads to robust results for compressed and low-resolution videos. The third part of the thesis is devoted to facial manipulation detection. We propose a novel approach, dubbed FaceMD, based on fusing three streams of convolutional neural networks to detect facial manipulation. The proposed FaceMD incorporates spatiotemporal information by fusing video frames, motion residuals, and 3D gradients to improve facial manipulation detection accuracy. We combine these three streams using different fusion methods and places to best use this spatiotemporal information, hence increasing detection performance. The experimental results show that the proposed FaceMD achieves state-of-the-art accuracy using two different facial manipulation data sets.