Research On Speech Perceptual Hashing Authentication Algorithm And Security Analysis Based On Compressed Domain

With the rapid development of Internet and multimedia processing technology, audio signals face the problems of illegal attack during transmission. As the most common form of audio, speech signals are widely used in daily communication, court evidence, and commands security. The authenticity and integrity verification of speech content are vitally important.A perceptual hash function is a kind of digital digest function in nature, which meets the basic security demand of unidirectionality. With the robustness for content-preserve operations and discrimination ability for malicious tampering, it has become a hot area of multimedia security such as image, audio and video.Speech signals are different from wideband audio in terms of signal characteristics, coding, transmission and storage ways. Current research on speech content identification aim primarily at raw wideband audio signals, but most proposed algorithms are unsuitable for speech and compressed signals. To solve the problems when applying traditional identification methods to the compressed speeches, this thesis proposes a new speech perceptual hashing algorithm for speech content identification with compressed domain based on MDCT spectrum entropy. The process begins by extracting the MDCT coefficients, which are the intermediately decoded results of compressed speeches in MP3format. In order to reduce the computational complexity, these coefficients are divided into sub-bands and the energy of MDCT spectrum is then calculated. Sub-bands of MDCT spectrum energy are then mapped to a similar mass function in information entropy theory. The function will be used as a perceptual feature and set to extract binary hash values. Experimental results show that the proposed algorithm keeps greater robustness to content-preserving operations while also maintaining efficiency. As a result of the partial decoding process, the real-time performance can meet the requirements of applications in communication terminals.To solve the problems of shortcoming in security and relevant analysis methods, this thesis proposes an audio watermark-based algorithm using entropy features in the compressed domain. A secret key is introduced to the process of hash extraction. The extracted hash values are performed with an Error Correction Coding function and used as authentication information, which will be embedded into the carrier speech signal. This method can save channel expense and protect the extracted hash values when they are transmitted in an unsafe channel. Experimental results show that the embedding process in the proposed method has no side effect on the perceptual content and maintains robustness to common attack. At last some tests are performed on the proposed algorithm using image security analysis methods. With the experimental results we come to the conclusion that the current security of audio perceptual hash is based on the secret key and special optimization for partial content tampering when constructing a perceptual hash algorithm.