DETECTION OF DEEPFAKE VIDEOS USING LONG DISTANCE ATTENTION

Abstract

Recent years have seen a significant advancement in deepfake methods, which has led to the creation of very convincing video content and serious security risks via facial video forgeries. Additionally, it is even more difficult and urgent to discover such phony films. The majority of detection techniques in use today approach the issue as a standard binary classification problem. Since the distinctions between fake and genuine faces are quite small, this work treats the topic as a specific fine-grained classification problem. A number of typical artifacts, such as generative faults in the spatial domain and inter-frame inconsistencies in the temporal domain, are shown to have been left behind by the majority of face forgeries techniques now in use. Additionally, a two-component spatial-temporal model is put out to capture, in a global viewpoint, the temporal and spatial forgery traces, respectively. A revolutionary long-distance attention mechanism is used in the design of the two components. Artifacts may be captured in a single frame using the one spatial domain component and in subsequent frames using the other temporal domain component. They produce patches that represent attention maps. With a more expansive perspective, the attention approach helps to better compile global data and extract local statistical data. Ultimately, similar to previous fine-grained categorization techniques, the attention maps are used to direct the network to concentrate on important facial features. The state-of-the-art performance of the suggested technique is shown by the experimental results on several public datasets, and the proposed long distance attention method can successfully capture crucial elements for face forging.