AbstractSteganography is the science of hiding information in some other medium. These media can be text, images, audio or video files. Steganographic analysis (steganalysis), on the other hand, is the science of detecting the existence of hidden information.
Many steganalysis methods have been introduced in the literature. These methods have been developed to combat specific steganography techniques and to detect data hidden in specific image formats. However, no single steganalysis method or tool can detect all types of steganography or support all available image formats.
One of the problems is there a need for more general system to cover different types of image formats and detecting wider range of stego images created by many steganography methods blindly. Blind steganalysis means detecting any stego image without knowing the type of the steganography method used or which type of file was embedded.
This thesis focuses on image steganalysis. A detection system is presented that combines three different steganalysis techniques. All technique address blind image steganalysis are rely on the extraction of selections of image features.
The first steganalysis technique presented here is based on extracting varieties of CGCM. The CGCM takes into account information of both colour correlations and gradients among the pixels in an image.
The second steganalysis technique works by extracting a number of histogram features. The features are extracted by exploiting the histogram of difference image, which is usually a generalised Gaussian distribution centered at 0. The histogram of difference image and the renormalized histogram are created for clean and stego images, there by using the peak value and renormalized histogram as features for classification.
Finally, the tested CGCM features and histogram features were merged together to improve the performance of the system. Merging two different types of features allow staking advantages of the beneficial properties of each in order to increase the system ability in terms of detection.
A large image database was created to train and test the system. The database included colour and grey images in various formats using both lossless and lossy compression.
The proposed detection system was trained and tested to distinguish stego images from clean ones using the Discriminant Analysis (DA) classification method and Multilayer Perceptron neural network (MLP). Stepwise Discriminate Analysis was applied in an attempt to find the best set of predictors. The Multilayer Perceptron (MLP) procedure was used to produce a predictive model for one or more dependent (target) variable based on the values of the predictor variables.
The experimental results prove that the proposed system possesses reliable detection ability and accuracy. The chosen classification methods show dissimilar performance in terms of classifying grey and colour images. The new system is a more generalized detector than previous systems, covering a wider variety of types of stego images, image formats and different hidden file sizes.
|Date of Award||Sep 2017|
|Supervisor||Saad Amin (Supervisor), John Filippas (Supervisor) & James Shuttleworth (Supervisor)|