Deepfake Detection Algorithm Based On Continual Learning

Deepfake technology can make the facial expressions in the video tampered with or the facial identity replaced,and the effect is so realistic that it is even difficult for the human eye to distinguish.This poses a huge threat to social and public information security.For the possible threats and impacts of deepfake videos,many deepfake detection models have been proposed,and some models have certain generalization.However,the generalization ability improved by these models is relatively limited.For possible future face swapping techniques,its generalization effect is limited.And the accuracy of these detection models is easily affected by adversarial examples and cannot be applied to real life.With the continuous emergence of more realistic deepfake generation methods,improving the generalization of deepfake detection methods is undoubtedly an extremely important research content.This paper intends to study the generalization of deepfake detection model by introducing the strategy of continuous learning,and mainly proposes the following two methods:1.A defense strategy for deepfake detection model based on bilateral filtering and adversarial training is proposed.From the perspective of passive defense and active defense,the bilateral filtering method and the joint adversarial training method are respectively introduced to defend the model.Bilateral filtering is applied in the preprocessing stage of the model without any modification to the model.Bilateral filtering is introduced to help denoise the input adversarial samples.The joint adversarial training method starts from the training phase of the model and mixes multiple adversarial samples and original samples to train the model.The introduction of joint adversarial training can train a model that defends against multiple adversarial attacks.The experimental results on the Face Forensic++ open-source dataset show that this strategy can effectively enhance the ability of the deepfake detection model to resist adversarial attacks,and has little effect on the original detection accuracy.2.A multi-color space fusion deepfake video detection model based on continuous learning is proposed to enhance the generalization of the model.The model focuses on the features after fusion of multiple color spaces.Different from the direct splicing of multiple color space features,a 1*1 convolution kernel is used to fuse the color features.This strategy is more lightweight and efficient.At the same time,the continuous learning algorithm MAS is introduced to combine with the deep neural network,so that the model can be generalized to the new deepfake generation algorithm.Experiments on multiple subsets of the Face Forensic++ open-source dataset show that the model has the best generalization performance for deepfake datasets in different domains.The detection accuracy is better than detection methods such as Xception Net and CLRNet.