Steganalysis Using LSB-Focused Statistical Features

In this paper, we present an image steganalysis model with a new texture feature set that is designed to take into consideration the pattern of embedding locations in a cover image. The chosen feature set in based on statistical texture features of images including gray level co-occurrence matrix (GLCM), Entropy, and additional statistical image features that can discriminate between clean and stego images. The guiding principle in choosing the feature set elements is that steganography techniques embed secret data in the right half byte of an image's bytes, the least significant bits, to avoid perceptible visual distortion that can result from embedding in the left half-bytes. Therefore, the proposed features are applied to 2-LSB, 3-LSB and 4-LSB bit planes of a cover image as well as the full-bytes. For the experimental work, the grayscale single channel image format was chosen for cover images, and we used the public BossBase1.01 dataset which consists of 10,000 PGM images. The selected classifier was the Support Vector Machine algorithm as implemented in MATLAB. Embedding of data in the cover images was based on 2LSB and 4LSB spatial domain schemes. The feature vectors of clean images, 2LSB stego images and 4LSB stego images, 10,000 each, were analyzed. The detection accuracy results of the validation phase was 99.41% for the combined clean and 4LSB images, and 99.02% for the clean and 2LSB stego images. The paper ends with conclusion and suggestions for applying the proposed model to multi-channel images, and for dealing with alternative steganography schemes.