Research On User Speech Data Privacy Protection

In recent years,smart voice terminals have been popularized in every corner of our lives.They bring a lot of convenience to our daily life,but also cause privacy hidden in users' speech to be leaked,which may be harmful to users' property and personal safety.In order to avoid the secret use of user speech by service providers and po-tential attackers,this paper proposes a cross-model voice emotion privacy protection framework mechanism Privacy++based on the cycle adversarial generative neural net-work.This mechanism can not only protect emotion privacy in users' speech,but also allow users continue to enjoy smart voice service,realizing an ideal trade-off between safety and convenience.Specifically,when a terminal device such as a smartphone re-ceives a voice command from the user,our mechanism converts the command from the original emotion to another emotion and protects the semantic information.Finally,ser-vice providers and attackers can not obtain the emotional information hidden in users'speech.In order to realize our mechanism,we firstly use resource-saving convolutional neural network and random forest emotion classifier respectively.These two classifica-tion algorithms are not only used to test the protection ability of our proposed mecha-nism for emotional privacy,but also uses the concept of supervised learning to help the Boosting algorithm enhance mechanism's privacy protection ability.Secondly,we de-signed our VC algorithm model based on CycleGAN,it realizes the conversion from the original emotion to the target,and at the same time protects the semantic information in users' speech to be consistent before and after the conversion,so it helps users to protect their privacy and enjoy the service provided by the smart terminal.In order to reduce the possibility of the mechanism being discovered and cracked by service providers and malicious attackers,we proposed the Rand om Choosing algorithm in the design of the mechanism.This algorithm is based on Monte Carlo sampling and the emotion distri-bution of raw speech.Algorithm result ensures that the overall emotional distribution of the speech after conversion is less different from the original distribution,and verifies the practicality of Privacy++.In order to verify the performance of the above algorithm,we conduct experiments from both objective and subjective perspectives.In the objective experiments,we use traditional and original evaluation metrics to verify Privacy++'s interference in accu-racy and precision for two classifiers in the edge-side and server-side.In the subjective experiment,we selected appropriate listeners to conduct mean opinion score test,and compared the scores of the filtered speech with the original speech in terms of emotion and content.The experimental results from two perspectives prove that our mechanism can maintain the content of the user's voice input and filter out the original sensitive emotion information,while not being restricted by the terminal's computing resource.