Deepfake Image Classification Algorithm Based On Deep Learning

In recent years,the deep synthesis technology,represented by the deep face forgery technology,has made remarkable success.Anyone can easily produce manipulated face images that are difficult to distinguish between real and fake with intelligent terminal devices.The malicious spread of these forged images will bring huge negative impact on media trust,social trust and political trust.It will put individuals and enterprises under reputation damage and property loss.More serverely,it may threaten social stability and national security.Most of face forgery detection algorithms process images in the spatial domain and extract local forgery clues through convolutional neural networks to distinguish the authenticity of the image.The progress of deep forgery technology based on the generative and adversarial networks is diminishing manipulated patterns in the spatial domain.And some postprocessing methods,such as consistency transformation and blurring,further reduce the appearance difference between real and forged face images.It’s getting much harder to detect those forged images under careful processing.Considering above problems,this thesis aims to improve the performance of face forgery detection from following two aspects.(1)In order to extract more discriminative forgery traces,this thesis proposes an algorithm base on frequency attention learning and contrastive learning.First of all,this thesis designs a frequency attention learning module,which focuses on the inherent frequency artifacts caused by the up-sampling operations.The frequency attention mechanism is proposed to enhance the forged information and suppress other unrelated regions.Additionally,a deepfake contrastive learning module is designed.It enlarges the distance between real and forged samples in the feature space and pulls real faces close.Meanwhile,the discrepancy between different forgery techniques is maintained.Finally,an interpretability reconstruction module is constructed to verify that the model concentrates on the manipulated regions.(2)In order to deal with the post-processing techniques applied on forged faces,such as blurring,compression,and noise,this thesis proposes an algorithm based on uncertainty learning.It adopts distribution estimation instead of point embedding estimation to improve the detection performance on fuzzy samples.In addition,this thesis designs a sampling strategy for difficult samples,which focuses on the top negative samples and trailing positive samples ranked by model scores,so as to achieve better overall ranking order.And improve the classification ability of the model.Based on deep learning technology,this thesis proposes multiple innovative methods for deep face forgery detection.Extensive experimental results indicate that the frequency attention learning and deepfake contrastive learning proposed in this thesis have achieved better performance in deep forgery dataset and cross-dataset experiments.At the same time,the uncertainty learning and difficult sampling strategy proposed in this thesis improve the classification accuracy of different baseline models.