Steganography is a technique that embeds data into multimedia such as photography, text, video. Among them, the method of embedding data into images is called Image Steganography. Image Steganography embeds data by manipulating pixel‘s value of image. Early stage of image steganography concentrated on not being founded in human''s eye and capacity of data, but that kind of methods leaved statistical traces, which can easily be detected. Later, steganographic technique has been progressed much elaborately and deftly, the steganographic images become more hard to detect.
The main purpose of steganography is to deliver data without detection to third parties. By embedding secret data into innocent image which called ‘cover image’, steganographer makes images which called ‘stego image’ that looks innocent.
Since Steganography can be exploited for secret information transmission, detection of Steganography is a very important problem. So Steganography and the method for detecting Steganography has been studied steadily.
The method for detecting Steganography is called Steganalysis. Along the growth of digital communication technology, Steganography and Steganalysis has been growed in competition. Steganalysis detects stego images by using differences between cover image and stego image. because of the tininess of steganographic traces, it’s relatively hard to detect stego image. In the early stages of steganalysis, stego image could be dectect by statistical differences. but along the progresses of digital technology, steganographic method became more sophisticated and had lesser traces. For detecting such sophisticate steganography, steganalysis methods use variety tools, like high pass filter, channel knowledge, machine learning and deep learning.
These days, convolution neural network(CNN) based deep learning technique show tremendous results in many computer vision’s area. Steganalysis method also adopt CNN structure for classifying cover image and stego image. In CNN based method, steganalysis shows high accuracy nearly 95%, and it’s almost saturated. Despite of the high accuracy, today’s steganalysis can only detect images with steganography. We can find images, but we can''t find data.
In this paper, as a first step toward extracting and restoring concealed data with steganography, we aim to detect the stego image and at the same time, classify which steganography algorithm is used. For this we propose CNN-based hierarchical Steganalysis structure, which has multiple networks for classifying N-stego algorithm. In general, noise in image by Steagnography is very tiny and the difference between Steganography has high similarity. The high similarity makes classification of Steganography algorithms harder. To improve the limits of a single CNN structure in classification of multiple Steganography algorithms problem, the proposed method classifies six classes of COVER, LSB, PVD, WOW, UNIWARD and MIPOD by combining existing CNN structures into hierarchical structures.
In this paper, multiple steganography algorithm classification results are shown through optimalizing CNN structure and hierarchical structure for steganalysis. The results of the experiment show the classification results for six classes through the data in Bossbase 1.01 of the hierarchical CNN structure and Our result shows the improvement of classifying accuracy through hierarchical CNN structure.