Tampered Image Detection using SVM Classifier

Abstract

"Steganography is the process of hiding confidential information in image such that contents of original image remain unaltered. Hence steganalysis algorithms used to detect such data embedding needs to be designed. In this work, features are designed to classify the given image as raw image (cover image) or image containing hidden data (stego image) embedded using LSB matching steganography algorithm. Finally, support vector machine classifier is trained using designed features. Two set of features are designed i.e one based on histogram of image and other based on information theoretic measure such as mutual information. Histogram of image is analyzed using short time Fourier transform and features based on centre of mass (COM) in frequency domain is designed. Statistical dependency between adjacent pixels in natural images is quantified using mutual information and novel features are designed based on that. Corel database containing 10,000 images is used for evaluating the proposed algorithm. Using this database, 20,000 images are made of same size out of which 10,000 are cover images and 10,000 are stego images. 85.71 % classification accuracy on the test set is obtained which is a significant improvement over the previously reported techniques."