On The Exploration Of Passive Forensics For Digital Image And Speech

Due to immense increase in data rate and development of novel data compression algorithms, digital multimedia (Video, Audio, Image etc) has become the best choice for information sharing. Growth in the usage of personal computer and multimedia software not only provides huge benefits to human race but also opens a back door for evil doers as it has become easy to edit, modify, tamper, and forge the multimedia data, according to the requirements of operator. This technology can be abused by malicious adversaries to generate the tampered or forgery copies for purposes such as pretending to be the owner of digital copy, removing/creating some objects in video, image etc. Recently, due to the spread of counterfeit, the integrity, reliability, and authenticity of multimedia data are the primary considerations of governments and as a result digital forensics has turned into one of the top research directions.According to the requirements of the integrity, reliability, and authenticity of multimedia data, this thesis is devoted to the exploration of passive digital forensics, including digital image steganalysis and digital audio forensics.Some pioneer research on the security analysis of reversible data hiding has been done in past, so three popular reversible data hiding methods were implemented. The statistical distortion introduced by data embedding was analyzed. Theoretical analysis illustrated that there was significant degradation in the horizontal component of statistical distribution of difference histogram. However, the vertical component of the statistical distribution of difference histogram was preserved. Quasi-Laplace Distribution (QLD), Generalized Gaussian distribution (GGD), and Adjacent Pixel Value (APV) model were introduced to describe the statistical distribution of difference histogram, separately. A simple and intuitive model, Horizontal and Vertical Difference Histogram (HVDH), which is based on similarity of horizontal and vertical difference histograms, was proposed to detect the presence of hidden message. Experimental results have shown that HVDH model can detect the presence of hidden information and it has outperformed QLD, GGD, and APV model too.Factually, only few blind steganalysis schemes for GIF image are present today. A novel blind steganalysis for GIF image based on generalized difference histogram and color correlogram was proposed. In the beginning, we introduced the general framework for hiding data in GIF image and analyzed the statistical distortion between the pixels. The generalized difference histogram of each RGB channel and the color correlogram was explored to capture the correlation between pixels. Finally, the absolute moments of characteristic function, Absolute Center of Mass (ACOM) of the characteristic function and high-order absolute moments of probability density function were extracted from the generalized difference histograms and color correlogram. Experimental results illuminated that the proposed scheme can not only detect the presence of hidden message embedded by several current steganography methods but also outperform Lyn’s blind steganalysis scheme, and Du’s target steganalysis scheme. It has comparable performance with Fridrich’s target steganalysis against steganography software EzStego. The proposed scheme has better capability of generalization as it is a blind steganalysis method.An Acoustic Environment Identification (AEI) scheme based on reverberation and background noise was proposed. Reverberation was considered as a kind of "intrinsic fingerprint" of environment which is used to capture the trace of environment. The features extracted from reverberation can be used for AEI. Firstly, blind de-reverberation method using block-based inverse filtering was introduced to separate the dry signal and reverberation. In order to improve the performance, background noise was taken as complementary feature. Background noise was estimated through particle filtering. Secondly,128-dimension feature vector, consisting of30Mel-frequency Cepstral Coefficients (MFCC),30Logarithmic Mel-spectral Coefficients (LMSE), mean, variance, kurtosis and skewness from reverberation and background noise were extracted for effectiveness verification. Experimental results on more than2240audio clips recorded in8different environments showed that the proposed framework was able to differentiate different environments, and outperformed the features extracted from the audio without de-reverberation and background noise estimation. The microphone independent test clarified that the performance of proposed features is not influenced by the type of microphone used to record the training and testing audio files. It has illustrated that the reverberation features is robust to MP3compression with various bit rates.