A Deepfake Detection Model Based on Xception and Self-Correcting Convolutional Networks

With the widespread use of AI technology, the cost of generating fake videos has decreased, and the realism of these fakes has significantly improved. This poses a serious threat to the authenticity of various fields such as social media, news dissemination, and film production. Furthermore, as video on many modern mobile devices is often compressed to reduce transmission bandwidth, low-quality fake videos are being widely circulated. Although detection techniques for high-quality videos have made progress, the technology for detecting fake low-quality videos remains lacking. Additionally, current fake detection methods perform well on single datasets, but their accuracy significantly drops when tested across different datasets. To address this issue, this paper proposes a deepfake detection model based on Xception and self-correcting convolutional networks. The model consists of a dual-branch structure: one branch is composed of a self-correcting convolutional network for extracting inconsistent feature information, while the other branch is based on an improved Xception network to capture deeper feature representations while reducing computational complexity and enhancing model efficiency. Finally, the extracted features are embedded into a Residual Network. To validate the effectiveness of the algorithm, the model was tested on the FaceForensics ++ and Celeb-DF datasets. Experimental results show that the proposed detection method outperforms other methods and demonstrates good detection performance across multiple datasets, further proving the effectiveness of the model.