Optimising acoustic features for source mobile device identification using spectral analysis techniques / Mehdi Jahanirad

Forensic techniques can be used to identify the source of a digital data. This is also known as forensic characterization, which means identifying the type of device, model, and other characteristics. Despite this, in recent years, the problem of multimedia source identification has extended its focus from identifying image/video sources toward audio sources. To determine the source of the audio several techniques have been developed. Those techniques work by identifying the acquisition device’s fingerprint as the detection features. However, the prior works have rarely considered audio evidence in a form of a recorded call. In filling that research gap, this thesis looks at intrinsic artifacts of both transmitting and receiving ends of a recorded call. Meanwhile, the influences such as speakers, environmental disturbances, channel distortions and noise contaminate the discrimination ability of the feature sets for source communication device identification. Hence, addressing robust feature extraction methods for source communication device identification is necessary. This study utilized spectral analysis techniques to investigate the use of linear and nonlinear systems for modeling the mobile device frequency response on the call recording signal. The context model allows computing the mobile device intrinsic fingerprints for the source mobile device identification. To achieve this aim, this study proposed a novel framework which extracts the mobile device intrinsic fingerprints from near-silent segments by using two spectral analysis approaches: (a) for linearized modeling, the proposed framework uses the cepstrum estimation technique and extracts entropy of Mel-frequency cepstral coefficients (MFCCs), (b) for non-linear modeling, the framework employs higher-order spectral analysis (HOSA) and utilizes the Zernike moments (ZMs) of the bicoherence magnitude and phase spectrum. Both models optimize acoustic features for source mobile device identification based on near-silent segments. The proposed feature sets along with selected feature extraction methods from the literature are analyzed and compared by using supervised learning techniques (i.e. support vector machines, nearest-neighbor, naïve Bayesian, neural network, logistic regression, and ensemble trees classifier), as well as unsupervised learning techniques (i.e. probabilistic-based and nearest-neighbor-based algorithms). The analysis was performed based on inter- and intra-model mobile device identification among 120 mobile devices in 12 models for speech and non-speech segments under different environmental influences, communication networks, and stationaries. For inter-model mobile device identification, the best performance was achieved with entropy-MFCC features and nearest-neighbor classifier, which resulted in an average accuracy of 99.63%. For intra-model mobile device identification, the best performance was achieved with ZMs of bicoherence magnitude and phase features and nearest-neighbor classifier, which resulted in an average accuracy of 98.45%.