Deep Fake Detection Based On 3D Geometric Detail Capture

Due to the rapid development of facial generation and forgery algorithms based on deep learning technology in recent years,although deep forgery technology can promote the development of film and television production,the social risks it brings are difficult to estimate.In the political field,"deep forgery" technology can trigger a crisis in the image of political figures.Therefore,adopting efficient and accurate methods to detect forged faces is of great significance for preventing internet risks.Most existing deep forgery detection technologies learn traces of forgery from subtle features between image pixel values,and there is a lack of further work on the interpretability of highfrequency detail features.In addition,there is still room for further exploration in extracting facial information from the perspective of 3D facial reconstruction.Based on the above issues,this article takes the extraction of interpretable forgery features from 3D facial reconstruction as the starting point,explores the possibility of the model in terms of accuracy and complexity,and verifies the generalization of the model:From the perspective of 3D facial reconstruction,this article introduces the newly released deep learning facial reconstruction technology DECA model in recent years.On the basis of deep forged facial data,the DECA model decomposes the true and false faces,and uses the displacement maps corresponding to facial photos as interpretable high-frequency geometric details,which can then serve as the feature basis for facial forgery detection.This article proposes a deep forgery face detection algorithm using 3D geometric details.This algorithm uses displacement maps extracted during the 3D facial reconstruction process,supplemented by cropped facial images,and uses convolutional neural networks to fuse and extract the features of the two.Then,a Transformer network structure is used for image classification to construct a geometric detail detection model.At the same time,the classification network structure was analyzed,and a more general network structure was used to simplify the Transformer network structure,achieving a balance between computational complexity and model accuracy.A training parameter of 21.4M,which was only half of the CVi T network model,was used,achieving a similar model accuracy rate of 86.72%.And the manifold distillation loss applicable to the Transformer module was applied in the training.Through the method of model distillation,the training effect of the model was maximized,and the training accuracy of the model was improved by 1.71%,with an AUC index of 86.42%on Celeb-DF.Under different compression rates of image quality,the geometric detail detection model improved the AUC index by 7.4% and 25.6% compared to the pixel noise feature detection algorithm Face X-ray at compression rates of c23 and c40,respectively.In cross dataset generalization experiments,the AUC index was achieved by 79.2% and 73.6% on the Celeb-DF and DFDC datasets,respectively,which were5.0% and 3.6% higher than the Face X-ray algorithm.A deep forgery spatiotemporal detection algorithm based on geometric textures is proposed.The algorithm uses UV texture maps obtained during the 3D face reconstruction process as input features,while still using displacement maps as input features.The geometric texture features are extracted using convolutional neural networks.Then,using the design principle of Vision Transformer,multiple consecutive frames of faces are input into the model together to capture temporal inconsistencies.Through experiments,it was shown that 86.46%,88.47%,and 88.18% accuracy rates were achieved on Face Forensics++,DFDC,and Celeb DF,respectively,which were3.23%,0.5%,and 0.98% higher than the 3D geometric detail model,further verifying accuracy and generalization.In summary,this article introduces a new approach to extracting forged trace features,and constructs single frame detection models and multi frame detection models.Numerous experiments have shown that this approach is effective in improving the accuracy and generalization of detection models.