Learning Meta Model for Strong Generalization Deepfake Detection
Dezhou Huang, Yuqing Zhang · 2024
Although deepfake technology is neutral, it can be maliciously used by criminals to cause serious security issues. These deepfake videos generated by deep learning technology are no different from real videos, posing a major threat to personal privacy and information credibility. Existing deepfake detection models face a core challenge: most models have limited generalization capabilities, and often have unsatisfactory detection results in the face of increasingly complex forgery technologies. To solve this problem, we introduce a two-stream deepfake detection model. One stream leverages the Video Swin Transformer to identify inter-frame discontinuities, a common anomaly in deepfakes. While another stream utilizes deep convolutional neural networks to detect facial texture inconsistencies, another telltale sign of a fake face. Furthermore, we use an improved meta-learning method called meta-learning for deepfake detection (MLDD) to train our model, which enhances the model’s adaptability and ability to quickly learn from multiple deepfake styles. Experimental results demonstrate that our model has superior performance and strong generalization compared to state-of-the-art existing techniques.