International Journal For Recent Development In Science And Technology

Deepfake techniques have reached an incredibly fast pace lately, making highly deceptive videos.Detecting such forging videos is urgently needed; it is also a rather challenging task.Unlike in the traditional approach, when this problem was treated more like a binary classification type, the approach here considered this problem as fine-grained classification due to the minute differences present between the real and the fake faces.Observations depict that most face forgery methods leave artifacts in spatial and temporal domains, that is, generative defects within individual frames and interframe inconsistencies.To catch these artifacts, a new spatial-temporal model, with two components designed through a long-distance attention mechanism, is proposed.The former component focuses on the detection of the artifacts in single frames and the latter one on capturing interframe inconsistencies.The attention mechanism produces patch-based attention maps that lead the network to pay attention to critical facial regions, enhancing the integration of global context and local statistical information.Experimental results on a variety of public datasets demonstrate that the proposed method achieves state-of-the-art performance.The long-distance attention mechanism effectively identifies pivotal forgery traces, providing a robust solution to the challenges of deepfake detection.