Research On Encrypted Speech Retrieval Method Based On Lightweight Neural Network And Perception Hashing

With the rapid development of cloud storage and Internet technology,multimedia data transmission and storage no longer become a bottleneck restricting the widespread popularity of speech communication.However,as an untrusted third-party platform,the cloud is extremely vulnerable to privacy leaks.How to ensure that users can efficiently implement encrypted speech retrieval from massive speech data under the premise of ensuring data security has attracted wide attention from many scholars.The key technologies involved in encrypted speech retrieval system,such as speech encryption algorithm,speech feature extraction,perceptual hashing scheme,and Lightweight neural network model construction are studied by using chaotic system theory,power normalized cepstrum coefficients(PNCC),speech perceptual hashing,and Mobile Net V3 neural network model.The main research work is as follows:1.In order to effectively protect speech data in the cloud and reduce the risk of plaintext leakage,a frequency domain speech encryption scheme based on a chaotic system is proposed.The program includes a speech encryption algorithm in discrete wavelet transform(DWT)domain based on Henon chaotic map,and another speech encryption algorithm in discrete cosine transform(DCT)domain based on hyperchaotic map.The experimental results show that the speech encrypted by the two encryption methods has a large enough key space,which can effectively resist brute force attacks.2.Aiming at the bad retrieval efficiency and accuracy of existing content-based speech retrieval methods,an encrypted speech retrieval method based on improved PNCC and perceptual hashing is proposed.Firstly,the raw speech is encrypted using the DWT domain speech encryption algorithm based on Henon chaotic mapping and uploaded to encrypted speech database in the cloud;Then,DWT and first-order difference coefficient were used to improve the PNCC feature extraction algorithm to extract speech features,and principal component analysis(PCA)was used to reduce the high-dimensional speech features to one dimension to obtain the frame features of the speech segment;Finally,the frame features are constructed as binary hash sequences using hash functions and uploaded to system hash index table in the cloud.The experimental results indicate that the speech perceptual hashing algorithm constructed by this method has good discrimination,robustness,and high retrieval accuracy and retrieval efficiency.3.Aiming at the problem that traditional neural network models have limited deployment on mobile devices due to their complex parameters,high requirements for device memory and computing power,the Mobile Nets V3 model is used in this thesis to propose an encrypted speech retrieval method based on the Bottleneck architecture Lightweight neural network model.First,the raw speech is encrypted using a DCT domain speech encryption algorithm based on hyperchaotic mapping and uploaded to encrypted speech database in the cloud;then,MFCC of the input raw speech is trained on the neural network,and the deep features of the fully connected layer output are combined with the perceptual hash algorithm to generate a binary hash sequence that can represent the speech,and uploads it to the system hash index table in the cloud;finally,the normalized Hamming distance algorithm is used for matching retrieval.The experimental results show that compared with the existing neural network,the network model generated by this method using the lightweight neural network configured with the Bottleneck architecture has fewer calculation parameters,and has higher retrieval accuracy and retrieval efficiency,and is beneficial to The model is deployed in embedded mobile devices with limited computing power and memory.