Tampering Detection And Steganalysis Of Digital Audio Signals

With the advancement in digital multimedia technologies, digital multimedia information is employed and spread widely, which has brought much convenience to information intercommunication of the people. Digital multimedia information can be easily modified, and spiteful tampering has resulted in serious trust crisis of multimedia information. On the other hand, as a kind of covert modification to digital multimedia information, steganography techniques may be utilized by terroristic and criminal organizations to disserve safety of the country and society. In recent years, tampering detection and steganalysis of digital multimedia information has attracted more and more attention.This thesis focuses on the techniques of digital audio tampering detection and steganalysis. Based on the theories and techniques of the audio processing, speech signal processing, speech compression coding and so on, detection of audio resampling, speech semantic tampering and speech timbre tampering, and compressed speech steganalysis is researched. The method systems of audio tampering detection and steganalysis are summarized and consummated. The contributions of this thesis are as follows:1. An audio tampering detection algorithm by detecting the existence of resampling processing is proposed. Based on analysis to spectral characteristics of un-interpolated and interpolated audio signals, a new spectral measure, Band-Partitioning Spectral Smoothness (BPSS),is extracted to scale the degree of spectral fluctuation. A comprehensive strategy is imposed to detect audio tampering with BPSS. By using the proposed algorithm, whether the audio signals are interpolated or stitched can be determined, and furthermore the original sampling rates of the interpolated signals can be estimated with high accuracy.2. A speech timbre tampering detection algorithm aiming at voice transformation is proposed. Voice transformation is the alteration to the sound characteristics of a person's voice while keeping the semantic information. Based on analysis to the basic model of speech production and voice transformation, vocal tract parameters and statistics of speech signals are extracted. Sensitive features are selected as classifying features by support vector machine recursive feature elimination method. Support vector machine classifiers are applied to get the decision of voice transformation detection. Speaker gender of transformed speech can also be detected with this algorithm.3. An algorithm for digital speech semantic tampering detection is proposed. This algorithm is based on speaking conditions analysis, which comprises background noise analysis and speaker condition analysis. Speech signals are divided into speech and silence segments. For silence segments containing noise, features in time domain and frequency domain are extracted for each frame; for each speech segment, rhythm and timbre features are extracted. Discontinuity points of the features of silence segments and speech segments are detected to get the tampering detection result. This algorithm can reduce the workload and subjective effect of manual detection as a tool of detecting and locating speech semantic tampering.4. For analysis-by-synthesis (AbS) compressed speech, a steganalysis method based on statistics of the pulse position parameters is proposed. Speech compressed by G.729 coding algorithm is taken as an example to analyze the characteristic differences of the pulse position parameters of cover and stego speech. In the steganalysis algorithm, the histogram flatness is measured, statistics including histogram flatness, the center of mass of histogram characteristic function, histogram variance, and the probability difference of 0 and 1, are used as distinguishing features, and SVM classifiers are employed. An improved algorithm is also proposed, in which local histogram features are used to improve detection accuracy. Experiment results show the proposed algorithms have good performance for steganalysis of speech compressed by G.729 algorithm. As to compressed speech coded by similar AbS-based coders, the proposed method is also applicable.