ATTENTION NETWORK FOR DEEP FAKE DETECTION

Abstract

The swift progression of machine learning techniques has led to the creation of "deepfakes," hyper-realistic synthetic media generated using algorithms like generative adversarial networks (GANs) and autoencoders. While the deepfake technology has beneficial applications in areas like entertainment and education, it poses serious significant risks such as misinformation, identity theft, and reputational damage. Therefore, the development of robust detection mechanisms, particularly those leveraging attention networks, is of paramount importance. This thesis focuses on the importance of attention networks for detecting deepfake, offering a comprehensive analysis of their effectiveness and challenges compared to traditional and contemporary methods. Attention networks, which enhance detection by focusing on critical regions of an image, are evaluated alongside convolutional neural networks (CNNs), multimodal detection techniques, adversarial training, transformers, and frequency-based models. Performance metrics like accuracy and the area under the receiver operating characteristic curve (AUC-ROC) are used to assessing models. The thesis emphasizes the significance of temporal coherence in video analysis and the role of frequency filters in identifying subtle artifacts. Attention-based methods are shown to offer superior performance in detecting fine-grained manipulations, achieving high accuracy and AUC-ROC scores. However, these models also face challenges related to computational complexity and generalization to novel deepfake techniques. The findings underscore the potential of attention networks to enhance deepfake detection, particularly in the real-world applications like social media moderation, news verification, and cybersecurity. This research not only helps advance but also helps in the understanding of deepfake detection, also bridges the gap between academic innovation and practical implementation. Future directions for research are suggested, focusing on improving computational efficiency, robustness, and ethical deployment of these technologies.