Research On Encrypted Speech Retrieval Method Based On Unsupervised Learning Hashing In Cloud Environment

With the rapid development of cloud storage and Internet technology,the demand for multimedia data sharing has increased significantly,and privacy leakage often occurs because of the rich semantic information in the speech.Content-based encrypted speech retrieval is a key technology for protecting user privacy and secure retrieval in application scenarios such as audio/speech retrieval,which is used to meet the demands of users for fast,secure and accurate retrieval of massive amounts of encrypted speech data in the cloud.Currently,existing supervised learning approaches require massive amounts of tagging information,which consume a lot of human and material resources.To solve the problem that supervised learning methods need to use tag information and it is difficult to obtain tags.This thesis mainly uses key technologies such as deep neural network model,unsupervised learning hash,index construction,B+ tree index,fully homomorphic encryption,and multi-threading to study the encrypted speech retrieval method based on unsupervised learning hash.The main research work is as follows.1.In order to solve the problems of low encryption and decryption efficiency of existing DGHV fully homomorphic encryption algorithm,large ciphertext expansion and only single-bit encryption form,a multi-threading based on speech DGHV fully homomorphic encryption scheme is proposed,and a many-to-one speech fully homomorphic encryption scheme model based on multi-threading is designed.Firstly,the pre-processing of speech data is completed and then the pre-processed speech data is split into binary strings by bits,and the cyclic encryption process is carried out by converting the split speech data into a matrix.Finally,the multithreading technique is used to encrypt the speech data concurrently.The experimental results show that the scheme has high security and the efficiency of encryption and decryption,and can resist various conventional attacks.2.To solve the problems that the existing encrypted speech retrieval methods rely on a large amount of human-annotated data for supervised learning,but it is difficult to obtain the speech data tagging information,and there are problems such as privacy leakage and low retrieval efficiency,an efficient and privacy-protecting unsupervised deep hash speech retrieval method is proposed.Firstly,an encrypted speech library is constructed and uploaded to the cloud server using the above speech encryption scheme.Then,the time series features of the speech are extracted and fed into the unsupervised deep hash learning framework Res Net18-GRU to learn the compact hash binary codes of the speech and establish a one-to-one mapping between the hash binary code and the corresponding encrypted speech.Finally,a pre-training mechanism is introduced to improve the training accuracy of the unsupervised learning network model.When users are retrieving,the normalized Hamming distance is used for retrieval matching.The experimental results show that the method has high retrieval accuracy,retrieval efficiency and security,and can be applied to the secure and efficient retrieval of massive unlabelled data.3.With the continuous growth of speech data in the cloud,it is difficult for the dynamic indexing strategy updates of existing speech retrieval systems to meet the real-time retrieval demands of users,as well as to avoid privacy leakage of speech data and improve retrieval accuracy and efficiency,an efficient encrypted speech retrieval method based on deep unsupervised hashing and B+ tree dynamic index is proposed.Firstly,the encrypted speech library is constructed using the above speech encryption scheme.Then,the compact binary codes of the learned speech are uploaded into the B+ tree index table using the above Res Net18-GRU network model,then a one-to-one mapping relationship is established between the hashed binary codes and the corresponding encrypted speech.Finally,the B+ tree index table is dynamically updated utilizing external B+ tree index techniques.The experimental results show that the proposed method achieves a retrieval accuracy of over 95.84%,greatly improving the retrieval efficiency compared with the retrieval method utilizing hash index tables,and effectively ensuring the security of speech data in the cloud.