Research On Few-shot Face Forgery Detection Algorithm Based On Meta-learning

In recent years,the technology of deepfake face manipulation has rapidly developed,allowing any user to create realistic fake faces to manipulate media content.While this technology can be applied to entertainment and filmmaking,the resulting fake face videos and images are often used to defame others and commit fraud,resulting in serious consequences.Therefore,designing a reliable deepfake detection algorithm has become a hot research topic in academia.However,there are currently numerous algorithms for creating fake face videos and images,and the differences between fake face videos generated by different algorithms are significant.Additionally,various manipulated videos are widely spread on social media,and are subjected to various compression and editing processes,making it difficult to obtain evidence for deepfake detection.Many existing deepfake detection algorithms perform poorly in detecting fake face videos in real-world scenarios and have poor generalization ability.In this paper,we address the common issues of poor detection performance on unknown algorithmgenerated samples and insufficient data compression resistance in existing deepfake detection algorithms,and we study the deepfake detection problem as a few-shot learning problem.We propose two different deepfake detection algorithms from the perspectives of domain-invariant features and data augmentation,respectively,and verify their performance in zero-shot and few-shot learning scenarios.Our main contributions in this paper are as follows:(1)We propose a multi-feature channel domain-weighted framework(MCW)based on meta-learning,which trains a detection model that can be generalized to unknown domains and can effectively detect compressed fake face videos.The MCW framework combines the RGB domain and frequency domain information of images using an improved meta-learning method,assigning element weights to the feature maps of the model in the channel domain,thereby improving the model’s generalization ability.Experimental results show that the MCW framework outperforms other compared algorithms in most scenarios,achieving the highest or second-highest performance.The MCW algorithm demonstrates its ability to generalize and resist compression in low-quality training images and across different generation algorithm scenarios,and it has better fine-tuning potential in few-shot learning scenarios compared to other algorithms.(2)We propose a metalearning-based method called the Meta-Feature Transformation(MFT)layers method,which learns the parameters of the feature enhancement layer through meta-learning and performs feature space-level enhancement on the data to simulate the distribution of image features,thereby improving the model’s cross-domain generalization performance.This algorithm performs various data augmentation operations in the highdimensional feature space,which can better expand the distribution of source domain data.Experimental results show that the overall performance of the MFT layers method is close to that of the MCW algorithm in(1),demonstrating the generalization ability and compression resistance of the MFT layers method.