Research And Implementation Of Deepfake Audio Detection Technology

With the development of deep learning theory and technology,deep neural networks have been widely used in speech recognition and voiceprint recognition and other scenarios,providing people with great convenience.However,deep learning technology also brings some negative effects.For example,a large number of fake audio can be quickly generated by speech synthesis technology or voice conversion technology,and the fake audio generated based on adversarial attack technology can easily deceive the target model and make it give wrong judgments.The fake audio generated by deep learning technology has an increasingly high degree of naturalness,and malicious attackers have begun to use this technology to conduct fraud and other illegal activities,which may cause serious economic losses to people.Therefore,there is an urgent need and great practical significance to study the detection technology of deep fake audio.The goal of this paper is to conduct technical research on deepfake audio detection and implement a prototype system for deepfake audio detection.This paper expands and introduces the source of deep fake audio,and divides it into two aspects: adversarial fake audio detection technology and non-adversarial fake audio detection technology.Specifically,the main work of this paper is as follows:(1)This paper designs and implements a real-time detection model of adversarial forged audio based on convolutional neural network,and proposes an adversarial audio sample generation technology based on adversarial attack technology,and generates an adversarial audio sample with 2400 adversarial forged audios.Fake audio detection dataset.Comparative experiments show that the model proposed in this paper has good performance on the adversarial audio detection task.(2)Based on the processing and analysis of the ASVspoof 2019 LA dataset,this paper proposes a model for non-adversarial forged audio detection that is robust to silent features.This paper proposes the conjecture that the silent feature of audio may seriously affect the performance of the non-adversarial forged audio detection model,and validates the conjecture through multiple comparative experiments.Experimental results show that existing models are sensitive to audio silence features,and the models may not have learned features that can effectively distinguish benign audio from fake audio.Therefore,this paper proposes a model for non-adversarial forged audio detection that is robust to silent features and achieves good results.(3)Based on the research of deep fake audio detection technology,this paper transforms,designs and implements the constructed model into a system platform that can provide services such as audio feature processing,fake audio generation,and fake audio detection.The actual use value of the work.