DETECTION OF DEEPFAKE VIDEOS USING LONG DISTANCE ATTENTION

Abstract

Deepfake videos pose a growing threat to digital trust, privacy, identity, and information security because manipulated facial content can appear visually authentic. Existing detection approaches based mainly on individual-frame analysis may overlook temporal inconsistencies and subtle artifacts distributed across distant video frames. This work presents a deepfake video detection framework that integrates video preprocessing, face detection and alignment, convolutional neural network (CNN) feature extraction, spatial-temporal analysis, long-distance attention, feature fusion, and binary deeplearning classification. Input videos are sampled into frames, normalized, enhanced, and analyzed for facial characteristics and temporal behavior, including motion, expression, lip synchronization, blinking, and head movement. The long-distance attention mechanism captures relationships between distant temporal regions and emphasizes informative forgery traces. The implemented system is developed using Python, TensorFlow, Keras, OpenCV, NumPy, Pandas, Matplotlib, and Scikit-learn. Evaluation uses accuracy, precision, recall, F1-score, and ROC analysis, while functional, integration, and acceptance testing assess system operation and robustness. Keywords:. Deepfake Detection, Video Forensics, Convolutional Neural Network, Spatial-Temporal Analysis, Long-Distance Attention.