Content-based encrypted speech retrieval scheme with deep hashing

In order to improve the limitations of manual features and poor feature semantics in the feature extraction process of existing content-based encrypted speech retrieval methods, and as well as improve retrieval accuracy and retrieval efficiency, a content-based encrypted speech retrieval scheme with deep hashing was proposed. Firstly, the original speech file is encrypted by using Henon mapping chaotic encryption to construct encrypted speech library. Secondly, adopting secondary feature extraction method to extract the spectrogram feature, and using the spectrogram as the input of the designed convolutional neural network (CNN) for model training and deep hashing feature learning, to obtain the deep hash binary code of original speech, and upload it to the deep hash index table in the cloud. In addition, the batch normalization (BN) method is introduced to improve robustness and generalization ability of the model. Finally, establish a one-to-one mapping relationship between the encrypt speech in the encrypted speech library and the hash sequence in the deep hash index table. When retrieving for speech users, the normalized Hamming distance algorithm is used for retrieve matching. The experimental results show that the deep hash binary code constructed by the proposed method has strong discriminability and robustness, and it still has high recall rate, precision rate and retrieval efficiency under various general content preserving operations.