Research On Voiceprint Biokey Generation Technology And Its Security Based On Deep Learning

With the progress of technology,the application of cloud computing and the popularity of cloud environments in recent years,the use of keys has gradually expanded and the forms of use have become more diversified.In order to overcome the shortcomings of traditional key memory,high memory cost,insufficient security and easy to be stolen,biometric key technology is widely used.Deep learning has strong capabilities in biometric recognition and other aspects,such as voice recognition and face recognition.Bio-key generation technology is a technique to generate high-strength keys using human biometric information,which includes voiceprint,fingerprint,face,iris,etc.Two models of voiceprint biometric key generation based on deep learning are proposed in this paper,and the security of one model is analyzed experimentally.This paper firstly proposes a dual neural network structure-based voiceprint biometric key generation model-MCP＿VP for text-related voiceprint patterns,which consists of three parts: voiceprint biometric key preprocessor,voiceprint biometric key stabilizer and voiceprint biometric key fuzzy extractor.The voiceprint key preprocessor combines traditional voiceprint preprocessing methods and deep learning to extract stable voiceprint features.In the voiceprint bio-key stabilizer,the instability between voiceprint samples can be effectively eliminated by using feature selection and deep neural networks to project the unstable features in multiple layers.The model generation strength is greater than 1024 bits,the accuracy rate is greater than 98.0%,and the false recognition rate is less than 1.5%.In order to further improve the generation speed of the voiceprint key generation model and reduce the false recognition rate,this paper designs a single deep neural network + feature vector selection + fuzzy extraction structure of voiceprint biometric key generation model-TSF for text-related voiceprint patterns.The Transformer+CNN structure can extract the feature information of the voiceprint more comprehensively and accurately,and the Sigmoid activation function can effectively binarize the output feature vectors for more accurate key generation.The accuracy of the voiceprint key generation of this model can reach more than 94%,while the false recognition rate can be kept below 0.0001.The generation strength of this model is greater than 1024 bits,which has high security.The TSF model uses the feature vector output from the deep neural network to generate the key,and whether the model can resist adversarial attacks determines whether the key is secure or not.This paper reproduces four algorithms of adversarial attacks against the TSF model,and then modifies the adversarial attack algorithm to update the adversarial samples by using the feature vectors output from the TSF model as labels.The adversarial samples generated by the original adversarial attack can make the model misclassify with a success rate of over 90%,but cannot make the model output the key of the specified voiceprint,and the adversarial samples generated by the modified adversarial attack algorithm cannot make the model output the key of the specified voiceprint.The adversarial attack experiment proves that the adversarial attack can make the TSF model misclassify,but the adversarial sample cannot generate the corresponding voiceprint key.