The rapid advancement of Deep Learning and generative models has led to the emergence of deepfake videos, which pose significant threats to digital security, privacy, and information authenticity. Deepfakes are synthetically generated media in which a person’s face or voice is manipulated using techniques such as Generative Adversarial Networks (GANs), making them highly realistic and difficult to detect. This project, “Detection of Deepfake Videos Using Long Distance Attention,” proposes an advanced deep learning-based framework that leverages Long Distance Attention mechanisms to identify subtle inconsistencies and temporal dependencies in manipulated videos. The proposed system utilizes a hybrid architecture combining Convolutional Neural Networks (CNNs) for spatial feature extraction and Attention-based models (such as Transformers) to capture long-range dependencies across video frames. By analyzing both spatial and temporal features, the model effectively detects anomalies such as unnatural facial movements, blinking irregularities, and inconsistencies in lighting or texture. The system incorporates preprocessing techniques including frame extraction, face detection, and normalization to enhance data quality. The model is trained on benchmark deepfake datasets and evaluated using performance metrics such as accuracy, precision, recall, F1-score, and ROC-AUC. The integration of Long Distance Attention enables the system to focus on critical regions and patterns across frames, improving detection accuracy compared to traditional CNN-based approaches. Additionally, the system can be deployed in real-time applications such as social media monitoring, digital forensics, and cybersecurity systems. This research contributes to the development of robust and scalable deepfake detection solutions, addressing the growing challenges of misinformation and digital manipulation.
ijesat (Sat,) studied this question.