Deepfake Detection Method Based On Supervised Machine Learning

Recently,with a series of achievements in artificial intelligence video synthesis and forgery technology represented by Deep Fake,we are entering a world where it is difficult to distinguish genuine and fake images.Four face manipulation technologies such as face synthesis,Face swap,facial attributes,and facial expression are widely used in the video consumer entertainment field,but at the same time,they are constantly engaged in false pornography,fake news,pranks,and financial fraud provide support.Various harmful applications of forgery technology have caused widespread social concern,and the Deepfake detection has become a hot issue in the academic world.In Deepfake detection,deep learning has grown the leading force in detecting fake images,and it is constantly playing an active role.However,forgery detection still has difficulties in offensive and defensive and actual combat detection.First,the circulation of forged images on the Internet will experience a lot of social media washing,which is not conducive to detecting potential forgery traces by the detector.So far,most data-driven Deepfake detection methods that use signal-level features have been severely affected by social media laundering.Secondly,there is a problem that the fake face detector does not perform well on images containing elements such as low resolution,blur,backlight,low light,and faces with complex lighting conditions.To enrich and improve the technology for detecting Deepfake,and through its role in increasing the cost of fraud by criminals,reducing the probability of network fraud,maintaining national public opinion,and political security,this article proposes two solutions to the above problems:First,we propose a Deepfake detection scheme that combines the Efficient Net B5 network and the attention mechanism to recognize fake faces.By adopting the attention mechanism to pay attention to the valuable feature content and feature location in the image,the model can adaptively obtain better network parameters and achieve end-to-end training.The experiment uses DFDC,one of the most comprehensive data sets for face forgery,and combines hue saturation,image compression,color dithering,scaling,and rotation to enhance data to train Efficient Net B5-Att.The experimental results show that the model trained by Efficient Net B5 using attention can obtain more positive information on the Deepfake,according to it to add additional supervision to bring better model performance;when the training set is different from the test set,the model is preprocessed.There is little difference in performance under different test sets.In the end,the experimental results prove that the fusion of the attention mechanism and the Efficient Net B5 network model can effectively improve the ability of Deepfake detection.Positive data enhancement significantly affects the models generalization ability and has higher adaptability to social media Laundering.Second,we propose a low-quality Deepfake detection method based on multi-scale feature extraction,make a detailed analysis of the low-quality Deepfake detection situation and focus on the low-quality Deepfake detection effect low-resolution insufficient image exposure.A fake face detection technology detection scheme based on HDR,image superdivision,and DCT technology feature extraction is proposed in poor situations.To better serve the model training,construct its data set in the experiment,train,and test the lowquality Deepfake detection model.Experimental results show that the network model trained with Efficient Net-RHD is applied to Deepfake detection,effectively improving low-quality Deepfake detection.At the same time,self-built data sets also play a positive role in lowquality Deepfake detection.In summary,this article studies the existing Deepfake detection algorithms,proposes practical solutions to the current detection models problems,enriches and perfects the algorithms in the fake detection field,and makes positive and beneficial contributions to the subsequent research Deepfake detection field.