Research On Information Hiding Security And Optimal Embedding Algorithms

Information hiding is the technique to embed secret information for coverttransmission, which is the very active and important component in the research ofinformation security field, and is the crossed subject involving information security,multimedia signal processing, pattern recognition and etc. It is the core and difficultyproblem to evaluate the security of information hiding system and to design thehigh-capacity, high-security and low-distortion embedding algorithms. This dissertationstudies the security of information hiding, the game between steganography and attackand the design of high-performance embedding algorithms. The main contributions canbe enurmerated as follows:First, the security measure method based on Chernoff information forsteganographic system is proposed. Through the steganalysis based on multi-samplehypothesis testing, the relationship between the union error probability and thedistributions of cover- and stego- signal is constructed, the Chernoff security definitionunder n-times observation is given, and the relationship between the error probabilitybound and security measure is analyzed.Second, the steganographic game model is proposed, which take the expectedsecure data transmission rateas the payoff function. According to the situations of therelative embedding rate choosing, the steganographic game is modeled as matrix game,Bayesian game and two-person zero-sum infinite game. We also give the equilibriumconditions and the corresponding equilibrium expected secure data transmission ratesunder difference situations. The rivalry realationship between steganography side anddetection side is modeled by game theory.Third, an embedding quantity estimation method based on potential function isproposed, and two estimate algorithm for LSB embedding and an anti-statistical-analysisLSB steganography are proposed. In the two methods, the histogram continuity of imagepixel values and the fitness bias of difference histogram are considerd as the potentialfunction, and some additional manners are used to improve the precision of estimation.The potential-function-based method provides a general design mind for embeddingquantity estimation. Fourth, the non-uniform quantization embedding method is proposed, and twohigh-performance embedding algorithm are presented to embed the data into image pixelprediction errors and image pixels. In the first algorithm, the neural network predictor isused to decrease the perceptual distortion. The second algorithm uses the structuralsimilarity disortion as the constraint, and takes the capacity as the optimal object torealize the optimal embeddin via dynamic programming. The research result shows thatthe non-uniform quantization embedding method can balance the inherent conflictingbetween capacity and perceptual quality operatively.Finally, the statistical properties preserving methods are studied at expense ofperceptual quality or capacity, and two secure embedding algorithms are proposed forJPEG images and vector-quantization-encoded images. In the steganographic algorithmfor JPEG images, the simulated-annealing algorithm is used to obtain the optimaladjustment vector to keep the statistics unchanged while expensing some perceptualquality. In the algorithm for VQ-encoded images, the statistical property is preserved byentropy decoding, and the perceptual quality is guaranteed by an optimal codebookpartition algorithm based on genetic algorithm. The optimization algorithms can be usedto banlance the inherent conflicting among security, capacity and perceptual quality.At last, the deficiencies in the dissertation are summarized, and some open issues ininformation hiding as well as the future work are presented.