Steganalysis of RGB Images Using Merged Statistical Features of Color Channels

This paper presents a steganalysis model that uses an enhanced grayscale statistical feature set, in the detection of data hiding in uncompressed RGB color images. A dataset of 3000 RGB images is created, using natural images from public sources, in TIFF and JPEG formats, that are converted to BMP format and resized to 512x512 pixels. The clean images are embedded with secret image data, using two payload schemes, 2 bits per channel (bpc) and 4 bits per channel. The selected feature set consists of 24 features per color channel, 72 features per image, which includes the Gray Level Co-Occurrence Matrix (GLCM) features, Entropy features, and statistical measures of variation. The feature set elements are calculated for individual channels, combined into image features vector. The steganalysis process is based on supervised machine learning, utilizing the Support Vector Machine (SVM) binary classifier's implementation in MATLAB. The results show very high detection accuracy for the two cases of 2-bpc and 4-bpc embedding schemes. Also, there are no noticeable differences in the detection accuracy between the two sources of images, even though un-compression of the JPEG images has reduced their noise contents. The paper ends with a conclusion and suggestions for future work.