SE-ResNet56: Robust Network Model for Deepfake Detection

AbstractIn recent years, high quality deepfake face images generated by Generative Adversarial Networks (GAN) technology have caused serious negative impacts in many fields. Traditional image forensics methods are unable to deal with deepfake that relies on powerful artificial intelligence technology. Most of the emerging deep learning-based deepfake detection methods have the problems of complex models and weak robustness. In this study, to reduce the number of network parameters, improve the detection accuracy and solve the problem of weak robustness of the detection algorithm, we propose a new lightweight network model SE-ResNet56 to detect fake face images generated by GAN. The proposed algorithm has high detection accuracy, strong robustness to content-preserving operations and geometric distortions, and strong generalization ability to different types of deepfake images generated by the same GAN.KeywordsDeepfakeGANLightweight networkStrong robustness