Research On Retrieval Method For Massive Encrypted Speech Based On Deep Learning

In multimedia data,speech has semantic function which makes it different from other sounds in nature,often contains more important information content and often closely related to personal and social privacy security.With the continuous development of cloud storage technology,this kind of security measures is particularly important.Therefore,how to protect the privacy of speech data and how to retrieve speech data efficiently has become a hot issue for researchers.This thesis mainly uses deep learning network model,deep hashing,speech signal processing and speech encryption technology to study the key technologies of encrypted speech retrieval system(speech feature extraction,deep hashing scheme construction,speech encryption).The main research work is as follows:1.In order to effectively avoid the risk of sensitive information leakage and ensure the security and privacy of speech data in the cloud,two 4D hyperchaotic speech encryption algorithms are proposed.The speech encryption algorithm based on 4D quadratic autonomous hyperchaotic system uses the generated four-dimensional chaotic sequences to perform inter-frame scrambling,intra-frame scrambling and diffuse operation.The 4D hyperchaotic speech encryption algorithm with quadratic nonlinearity can perform inter-frame and intra-frame scrambling at the same time,which can reduce the encryption time.The experimental results show that the two speech encryption algorithms have high key space and can effectively resist exhaustive attack.2.In order to overcome the feature extraction defect of existing content-based encrypted speech retrieval methods,and solve the problem of low retrieval accuracy caused by high dimensional and temporality of speech data.An encrypted speech retrieval method based on convolution neural network(CNN)and deep hashing is proposed.Firstly,4D hyperchaotic encryption algorithm is used to generate encrypted speech library,and then deep hashing technology is used to generate deep hashing sequence on Log Mel Spectrogram and MFCC features to achieve efficient retrieval.The experimental results show that the proposed method has good discrimination and robustness to amplitude change compared.Meanwhile,the proposed method has high recall,precision and retrieval efficiency after various content preserving operations.3.In order to achieves efficient retrieval of massive encrypted speech in cloud environment and improve retrieval accuracy and retrieval efficiency,an encrypted speech retrieval scheme based on long short-term memory(LSTM)neural network and deep hashing is proposed.This scheme mainly uses LSTM network to extract deep semantic features of speech and combining with hash function to generate deep hashing codes.The normalized hamming distance is used to achieve retrieval matching,which improves the efficiency and accuracy of massive speech retrieval.The experimental results show that the algorithm has good discrimination,robustnes,recall,precision and retrieval efficiency.4.For speech deep feature extraction,CNN can only extract local features.And the problem that LSTM has a large amount of learning calculation,long processing time,and the loss of information becomes very obvious as the length of speech increases.Therefore,an encrypted speech retrieval method based on CNN-BiLSTM and deep hashing is proposed.The method first extracts the Log Mel Spectrogram and MFCC features of the original speech,then enters CNN and BiLSTM networks in order to perform model training and deep feature learning,and generates a deep hashing sequence to achieve retrieval matching.The experimental results show that the method has good discrimination,robustness,recall and precision compared with the existing methods,and has good retrieval efficiency and accuracy for long speech.