Research On Generalizable Deep Learning-Based Face Forgery De Tection Method

In recent years,research related to deep learning has developed rapidly,and a series of fake face generation algorithms based on deepfake technology have begun to enter the public’s view.A large number of fake videos based on deepfake technology have appeared on major internet media platforms,which pose a serious threat to personal privacy,public trust,social stability,and national security.Therefore,researching deepfake detection has become an important topic.Currently,there are many deepfake detection algorithms that have analyzed and detected deepfake results from different perspectives and achieved good results on different datasets.However,most detection algorithms exhibit deficiencies in generalization when faced with cross-dataset challenges.The rapid development of deepfake technology and the complexity and diversity of deepfake methods have raised higher requirements for deepfake detection.Meanwhile,Enhancing the capability of detection algorithms to generalize within the current dataset is a key obstacle in deepfake detection.In response to the above research problems,this study mainly focuses on the generalization problem of deepfake detection,aiming to improve the cross-dataset detection performance of the model while ensuring stable detection performance within the dataset.This article designs a deepfake detection framework and improves it based on the deepfake detection task,including:(1)This study proposes a multi-layer deepfake detection framework based on image decomposition and inter-layer spatial attention transfer.Compared with directly using pre-trained models,this study attempts to extract fake features by designing a network model that can extract differences between real and fake faces.This study adopts the idea of decomposing the original input and extracting features layer by layer,using Laplacian pyramid for image decomposition to separate the differences and commonalities between real and fake samples as much as possible.To enhance the generalization ability of the deepfake detection neural network,the inter-layer spatial attention transfer module is utilized to improve the model’s ability to extract differences.In the experimental part,our results show that the model proposed in this study performs well within the dataset and has strong performance in cross-dataset testing.(2)This study proposes a feature optimization for deepfake detection based on central difference convolution and supervised contrastive loss.Based on the solution proposed in(1),we optimize the input layer and feature output layer of the model.To fully utilize the data characteristics of the deepfake detection task and address the sparsity residuals and fixed decomposition problems that may exist in Laplacian image pyramids,we design a relatively flexible and trainable image decomposition module based on central difference convolution.To avoid the degradation of the decomposition process,we also design a decoder to reconstruct the remaining features.In addition,as the dataset contains various fake algorithms,and the fake traces between algorithms are different,we introduce supervised contrastive loss to constrain the feature space of the high-level branches of the model to optimize the extracted feature distribution and boost our deepfake detection model’s generalization capability.In the experimental part,our results show that after introducing central difference convolution decomposition and supervised contrastive loss,the model extracts richer features,and the feature distribution is more in line with the requirements of the detection task.The performance of this study is stable within the dataset and excellent in cross-dataset testing.