Research On Deepfake Speech Detection Based On Improved Mimic Defense

With the rapid development of voice technology,deepfake voice has become more and more convenient.This technology has gradually moved from a high threshold to the public,turning creative deepfake voice into public entertainment.However,while deepfake voice is convenient for the public,it also confuses the public's audio-visual.The technology of deepfake voice based on deep learning makes the real news no longer true,which makes the important audio-visual news difficult to be believed,affects the public's life,and even causes property loss and threat to life safety.Therefore,the detection of deepfake voice becomes particularly important.First,a variety of methods to enhance the wavelet packet energy spectrum and integrate speech feature extraction are proposed to improve the deepfake speech detection mimic architecture.In the detection process of deepfake speech,an improved mimic defense structure is added.After multiple features are extracted from the speech by strengthening the wavelet packet transform and other methods,the number of features is randomly selected to generate a sub-feature data set,and the sub-feature dataset is generated through the GRU model and SVM model.Perform pre-training with the differential Transformer model,and screen out the pre-training model with better effect.In addition,using the dissimilar redundancy structure,the pre-trained GRU model,SVM model and differential Transformer model are used to construct heterogeneous redundancy.The final execution model is determined by random selection.The deepfake speech detection method based on improved mimic defense improves the security of the system and can detect deepfake speech efficiently and safely.Secondly,a deepfake speech detection method based on differential Transformer encoder is proposed,which improves the upper limit of the system's detection effect on deepfake voice.Since there are obvious differences in the low frequency,medium frequency and high frequency in the feature,the uneven feature information can be converted into stable information through the difference operation,so as to obtain the dynamic characteristics of speech signal.The experimental results show that the deepfake speech detection method based on the differential Transformer encoder can effectively capture the dynamic difference characteristics of audio signals,improve the detection effect of deepfake speech,and effectively improve the upper limit of detection performance of the improved mimic architecture system.To sum up,the deepfake speech detection method based on improved mimic defense proposed in this paper is an efficient and secure detection method.By enhancing the energy spectrum features of the input speech wavelet packets and extracting the speech features together with Mel cepstral coefficients(MFCC),the speech features are input into the redundant heterogeneous model structure based on the differential Transformer model,the GRU model and the SVM model.After the comparison of multiple sets of experiments,the experimental results show that the effect of deep forgery speech detection based on improved mimic defense is significantly better than that of traditional feature extraction methods and traditional models,and the security is better improved,which can effectively prevent generative adversarial network attacks.