Generalized Meta-Learning EfficientNet Vision Transformer Model for Deepfake Detection

Digitally manipulated images that are realistic-looking but fake, which are known as Deepfake. With the remarkable developments in deep generative models, the accessibility and accuracy of manipulated technologies are increasing, leading to fake videos becoming increasingly difficult to identify. Different facial forgery techniques result in complicated data distributions, but Deepfake detection techniques based on CNN(convolutional neural network) architecture are utilized in the majority of Deepfake detection models as binary classification problems. In this paper, we propose a model, named MEViT, which uses a combination of EfficientNet Vision Transformer with a meta-learning-based technique to improve the generalization of the detection model. Furthermore, we propose a learning process to update the model and introduce pair-discrimination loss and domain adjustment loss to improve detection ability across various domains. We also create various experiments on several Deepfake datasets and compare our proposal with many state-of-the-art works to prove the efficiency of our approach.