A retrieval algorithm for encrypted speech based on convolutional neural network and deep hashing

In this paper, we propose a retrieval algorithm for encrypted speech based on the convolution neural network (CNN) and deep hashing. It is used to overcome the feature extraction defects of the existing content-based encrypted speech retrieval methods, and solve the problem of low retrieval accuracy...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Multimedia tools and applications Ročník 80; číslo 1; s. 1201 - 1221
Hlavní autoři: Zhang, Qiu-yu, Li, Yu-zhou, Hu, Ying-jie
Médium: Journal Article
Jazyk:angličtina
Vydáno: New York Springer US 01.01.2021
Springer Nature B.V
Témata:
Algorithms
Artificial neural networks
Audio data
Computer Communication Networks
Computer Science
Convolution
Data Structures and Information Theory
Encryption
Feature extraction
Hash based algorithms
Multimedia Information Systems
Neural networks
Retrieval
Special Purpose and Application-Based Systems
Speech
Training
Convolutional neural network (CNN)
Batch normalization algorithm
Encrypted speech retrieval
Deep hashing
Speech feature extraction
ISSN:1380-7501, 1573-7721
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:In this paper, we propose a retrieval algorithm for encrypted speech based on the convolution neural network (CNN) and deep hashing. It is used to overcome the feature extraction defects of the existing content-based encrypted speech retrieval methods, and solve the problem of low retrieval accuracy caused by high dimensional and temporality of audio data. Firstly, the study encrypts the original speech by the three-dimensional chaotic encryption algorithm and uploads it to the encryption speech library in the cloud. Since CNN can well capture the basic semantic structure features of speech data, we use CNN as a feature extractor to extract deep features from Log-Mel spectrogram/MFCC. The batch normalization algorithm is introduced in the training process, which improves the speed of network fitting, reduces the training time, and improves the retrieval efficiency of the system. Secondly, the deep features extracted from CNN are combined with the hash function to construct the system hashing index table. Finally, the retrieval is implemented by the normalized Hamming distance algorithm. The experimental results show that the proposed algorithm has better discrimination, robustness to amplitude change compared with the existing methods. Meanwhile, the proposed algorithm has a high recall, precision, and retrieval efficiency after various content preserving operations.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1380-7501
1573-7721
DOI:10.1007/s11042-020-09748-y