Deep Learning-based Forgery Detection Research

With the advent of Generative Adversarial Networks(GANs),face forgery technology has been extensively exploited,necessitating accurate detection of forged faces.However,forgery techniques continue to evolve,akin to the competition between generators and discriminators in GANs.To address this challenge,this study delves into several key aspects of forgery detection,including biometric signal-based detection and the impact of content information,adopting a progressively refined approach,ranging from simplistic to intricate methods.Initially inspired by Remote Photoplethysmography(rPPG),PPG signals correspond to the periodic variation of skin color in facial videos caused by heartbeats.We discovered that,despite the inevitable loss of PPG signals during the forgery process,a hybrid PPG signal persists in forged videos,exhibiting a unique rhythmic pattern dependent on the generation technique.Based on this pivotal observation,we propose a face forgery detection and classification framework comprising:1)a Spatio-Temporal Filtering Network(STFNet)for PPG signal filtration;2)a Spatio-Temporal Interaction Network(STINet)for PPG signal constraint and interaction.Furthermore,through a thorough analysis of the forgery generation process,we introduce intra-source and inter-source fusion strategies to enhance the framework’s performance.A wealth of empirical evidence attests to the superiority of our approach.However,while these convolutional neural network-based face forgery detection techniques demonstrate significant achievements during training,they struggle to maintain comparable performance during testing.We found that detectors tend to focus more on content information rather than artificial traces,indicating a heightened sensitivity to inherent dataset biases,resulting in severe overfitting.Prompted by this insight,we designed an easily embeddable content information removal disentanglement framework and further proposed a Content Consistency Constraint(C2C)and a Global Representation Contrast Constraint(GRCC)to bolster the independence of disentangled features.Additionally,we ingeniously constructed two imbalanced datasets to investigate the impact of content bias.Numerous visualizations and experiments demonstrate that our framework not only disregards content information interference but also guides detectors in unearthing suspicious artificial traces,achieving competitive performance.In conclusion,this study provides robust theoretical underpinnings and practical guidance for achieving precise and stable forgery face detection.Extensive experimental results validate the superiority of our method in addressing the ongoing advancements in forgery techniques and inherent dataset biases,laying a solid foundation for future research.