Design and Development of Machine Learning Based Framework for Video Forgery Detection

Abstract

The rapid spread of digital video content with the advancements in editing tools has newlinemade it challenging to detect forgeries in video. This thesis presents a sequence of newlineinnovative methods to address different types of video forgery, with a focus on intraand#65534;frame and inter-frame inconsistencies, as well as deepfake detection. newlineThe first approach considered here is the detection of intra-frame copy-move forgery. newlineThe technique identifies manipulated frames where regions are duplicated within a newlinesingle frame in an attempt to alter the content. This method has been documented newlineextensively based on foundational concepts and experimental analysis, employing the newlineREWIND dataset that demonstrates robust capacity in detecting copy-move forgeries. newlineThe second proposed method applies a hybrid deep-learning model for deepfake newlinedetection using transfer learning applied with architectures based on GRU. This newlinemethod applies the capabilities of GRUs to learn unique temporal patterns of newlinedeepfakes validated in sets such as FaceForensics++, CelebDF, and DFDC, showing newlinehigh accuracy in detecting subtle manipulations both in spatial and temporal domains. newlineThe third one follows hybrid modeling to take the CNN and RNN models and newlineintegrate them with optical flow analysis. Such motion-based identification of newlinedisparities from frames improves the ability to find advanced video forgeries. Indeed, newlineit is by bringing these ideas of optical flow and deep learning into this method that it newlinemay be used to accurately identify complex forgeries and contribute to a strengthening newlineframework in the preservation of digital authenticity. newlineThe final technique is an advanced method for detecting video forgeries, designed to newlineenhance efficiency, It starts with steps like frame extraction and alignment, followed newlineby extracting important features and classifying them. This multi-step approach helps newlinexi newlinethe method maintain good quality data and high performance, leading to accurate and newlineefficient detection.