Research On Learning Based Detection Of Image Hidden Message

With the rapid development of network communication technology, digital multimedia data can be quickly transmitted all over the world, which has brought much attention to the communication security. Steganography is the art of covert communication by invisibly embedding message into innocuous looking multimedia data, which can be used to transfer secret messages including national confidentiality, commercial information and privacy etc. However, as steganographic technologies are used to serve people, it can be also utilized by the illegal organizations to transmit the injurious information, which harms the social security and stabilization. Therefore, the detection of image hidden message, which is an opponent to steganography, possesses great significance.In this thesis, we concentrate on learning based detection of image hidden message. Features which are sensitive to the message embedding are extracted from large image database to train the classifier which then is utilized to detect the hidden message. Feature extraction plays a key role in this method. The main contributions of thesis are presented as follows.(1) Spatial LSB matching can be modeled as adding independent additive noise to the image, which will smooth the image histogram and disturb the correlation among image pixels. Accordingly, the image histogram gradient energy is calculated as histogram feature, and the neighbourhood degree histogram and the run-length histogram are designed to extract features about image correlation. Wavelet transform is utilized to denoise the test image so as to generate a calibrated image. Then, the features are extracted from both test and calibrated features, and the ratio of corresponding features is utilized as final features. Support vector machine (SVM) classifiers are trained and tested based on the proposed features with two large image databases. Experimental results testify that the proposed method possesses reliable detection ability and outperforms the two previous state-of-the-art methods. Further more, the performance of histogram and correlation features in this thesis are compared, which reveal that the two types of features have their own advantages in the detection of non-compressed and JPEG-compressed images respectively.(2) The detection performance of classifier could be enhanced by appropriately increasing the dimensionality of feature vector. The above-mentioned features are extended to increase the detection accuracy. Besides the histogram gradient energy, the absolute differences between adjacent elements of image histogram are calculated as the histogram features. Four difference images are constructed from horizontal, vertical, main diagonal and mirror diagonal directions, from which co-occurrence matrix is utilized to extract features which are essentially based on image correlation. A calibrated image is generated by embedding message into the test image. Features are extracted from both test and calibrated images; and the ratios of corresponding features between test and calibrated images are calculated as the final features. SVM are utilized to train the classifiers on a JPEG-compressed and a non-compressed image databases. Experimental results show that the proposed method holds a better performance,(3) Steganographic methods that use JPEG images as covers usually modify the quantized DCT coefficients of images so as to embed the message bits. Therefore, the features extracted from DCT domain are expected to be the most sensitive to the embedding. Three kinds of intrinsic characteristic of DCT coefficient, including distribution of coefficients, intrablock correlation and interblock correlation, are analyzed and exploited to extract features. Firstly, histogram is utilized to model the distribution of coefficients in order to extract the histogram features. Secondly, the difference is explored to express the intrablock and interblock correlation of DCT coefficients, and then co-occurrence model is used to extract corresponding features. Finally, the three kinds of features are calibrated by "decompress-cut-compress" method. SVM is utilized to learn and discriminate the differences of features between original and stego images based on a large image database. Experimental results demonstrate that the three feature sets individually succeed in attacking the four typical steganographic tools to some extent, with the intrablock feature set performing the best. Furthermore, the comparison experiments show that the jointed feature set not only outperforms the three individual feature sets but also surpass a previous state-of-the-art steganalysis method.(4) The existing steganalysis methods extract features with increasing dimensionality in order to enhance the detection accuracy. However, with more features, the computational complexity increase and the effect of the new extracted feature is obscure. In this paper, a feature selection method based on genetic algorithm is proposed. Fitness function is constructed with support vector machine; and the parameters of crossover operator and mutation operator are also suitably configured. In addition, similarity of individuals (SI) in each generation and transformation between two adjacent generations (TG) are calculated to judge whether the algorithm is trapped in a local area. Then, the algorithm can be released from the local without satisfying solution. Experimental result shows that the classifiers trained with the feature subset hold better accuracy and speed than that trained with feature universal set.