Search

Video compression is pervasive in digital society. With rising usage of deep convolutional neural networks (CNNs) in the fields of computer vision, video analysis and video tampering detection, it is important to investigate how patterns invisible to human eyes may be influencing modern computer vision techniques and how they can be used advantageously. This work thoroughly explores how video compression influences accuracy of CNNs and shows how optimal performance is achieved when compression levels in the training set closely match those of the test set. A novel method is then developed, using CNNs, to derive compression features directly from the pixels of video frames. It is then shown that these features can be readily used to detect inauthentic video content with good accuracy across multiple different video tampering techniques. Moreover, the ability to explain these features allows predictions to be made about their effectiveness against future tampering methods. The problem is motivated with a novel investigation into recent video manipulation methods, which shows that there is a consistent drive to produce convincing, photorealistic, manipulated or synthetic video. Humans, blind to the presence of video tampering, are also blind to the type of tampering. New detection techniques are required and, in order to compensate for human limitations, they should be broadly applicable to multiple tampering types. This thesis details the steps necessary to develop and evaluate such techniques.