An Information Theoretic Image Steganalysis for LSB Steganography

Steganography hides the data within a media file in a subtle way while steganalysis exposes steganography by using detection measures. Traditionally, steganalysis reveals steganography by targeting perceptible and statistical properties of the image for improving the security of steganography methods. In this work, we target LSB image steganography methods by using information theoretic metrics for steganalysis. Our technique works in two phases. First, the features of embedded image are extracted, and then this image is analyzed on the basis of entropy and joint entropy features corresponding to the original image. Second, from these extracted features of the image, we train SVM and ensemble classifiers. The classifiers discriminate cover image from the stego image. For evaluating the robustness of our method, several attacks over the stego images are applied. These attacked stego images are then analyzed for the detection reliability of our method. Original images and LSB embedded images of the dataset are analyzed by comparing information gain from entropy and joint entropy metrics. Results suggest that entropy of the stego images are more preserving than that of joint entropy. Experimental results show that before histogram attack, detection rate with entropy and joint entropy are 70% and 98%, respectively while after the attack entropy metric gives 30% detection rate and joint entropy gives 93% detection rate. Therefore, joint entropy proves to be a good steganalysis measure having fewer false alarms for across a range of hiding ratios.