A Study on an Efficient Tampering Detection and Localization Method for Speech Signals

Abstract

Due to the development of digital technologies, new social issues relating to malicious attacks and unauthorized tampering to speech, such as content replacement and voice morphing, have arisen. Using advanced speech analysis and synthesis tools enables the speech to be tampered without leaving any perceptual clues. As an important information carrier, the originality, integrity, and authenticity of speech signals should be strictly confirmed. Authentication and tampering detection of digital signals is one of the main applications of digital watermarking. In this thesis, an efficient speech watermarking method is proposed to generate self-embedding speech signals in that hash representation of a speech signal, which is assumed as watermark, is embedded into the signal itself without affecting the original quality. The proposed system is intended to satisfy blindness, inaudibility, and fragility against malicious modifications. The proposed watermarking method in this thesis is a kind of fragile watermarking and thus the hash information (watermark) in the tampered regions is destroyed when tampering occurs. This feature helps the receiver to detect and localize the tampering regions by comparing the original hash information and the extracted hash from the received speech. The perfect match of the hash information confirms the integrity and originality of the received speech; otherwise it indicates tampering. In this thesis, performance of the proposed system is tested on 40 read speech files which are International news and Burmese news read by 20 female and 20 male announcers. The proposed method is implemented in MATLAB and fragility against malicious modification is evaluated by applying various kinds of tampering such as compression, zeroing, adding noise, time scaling, and reverberation attacks on the watermarked speech files. Experimental results show that the proposed method is relevant with the main requirements of a good watermarking scheme: inaudibility and blindness in addition to fragility. Therefore, the proposed system is really useful for applications of criminal investigation and digital forensics where the integrity and originality of the speech evidence is extremely important.