Study On Retrieval For Encrypted Speech And Recovery Watermarking-Based Speech Authentication

Nowadays computers and communication satellites have been gradually spread around the world, and modern technologies such as information technology, the Internet, telecommu nications have rapidly developed, so more and more people benefit. However, the amount of storage and computing architecture of the Internet has been unable to meet the large multimedia data management and computing requirements, then the cloud technology is turning up and provides a new idea for multimedia services. Retrieval and copyright protection for multimedia information of vast data in the cloud platforms have broader concerns now.In this paper, a perceptual hashing-based retrieval scheme for encrypted speech is designed. In addition, a speech content-based recovery watermarking scheme for speech authentication is designed. The main work is as follows:(1) First, a speech scrambling encryption scheme based on AES is used, and then put forward an encrypted speech retrieval algorithm based on perceptual hash which is generated by the characteristics of the speech itself. The algorithm will extract the pitch as speech characteristics, and generate the perceptual hash, then embed it in the encrypted speech by using the method of LSB. When a user sends a retrieval request, the administrator can get the result by computing the normalized Hamming distance of the perceptual hash between the target speech and the extracted one. There’s no need to decrypt the encrypted speech during the process. Experimental results prove that the discriminability and uniqueness of the proposed scheme are very well, and the robustness to deal with common speech processing is also strong. In addition, the ideal recall ratio and precision ratio are also obtained.(2)A recovery watermarking algorithm for speech content authentication is designed. In this algorithm, the watermark is generated by the speech itself. First, a one-to-one mapping relationship for each speech frame is established, and then the characteristic of different length is extracted according to the judgment of whether the speech sub-band is plain, and is embedded into the samples with the mapped frame as a watermark. In the embedding process,12bits synchronization code and frame number is added into the watermark. When a user needs authenticate the contents, the administrator can compare the extracted watermark and the feature value which generated based on the speech to judge whether the speech has been attacked. The watermark extracted from the mapping frame can be used to recover the speech. Experimental results show that the proposed algorithm is inaudible. It is not only sensitive against additive noise, re-quantization, etc., but also sensitive against attacks like mute, replace, cut, insert, etc. Especially, it can detect and localize the tampered regions effectively, and it can recover the speech from the tampered regions.