Simple and effective approach for construction of universal blind steganalyzer

Abstract

The goal of universal blind steganalysis is to detect all known (already existing) and unknown (previously unseen) steganographic algorithms without knowledge of the exact stego algorithm used by the steganographer. However, a binary blind steganalyzer trained on cover images and stego images randomly selected from “known stego images” (i.e., stego images produced by multiple “known” stego methods with a mixture of payloads), may fail catastrophically on unknown stego methods although shows superior performance on known stego methods. Additionally, unsupervised outlier detection and one-class classification approaches are less likely to fail to detect unknown stego methods but yield high false positive rates. Motivated by these observations, we explore a simple and effective approach for construction of universal blind steganalyzer to achieve overall good performance on both known and unknown stego algorithms. First, we compute Local Outlier Factor (LOF) scores of known stego sample points (feature vectors) with respect to test sample points. Then, we choose stego images with the lowest LOF scores from known stego images as training stego images. Finally, we train a binary classifier on cover images and chosen training stego images for test. Experimental results confirm that the proposed approach performs significantly better than the random sampling-based binary classification method, unsupervised outlier detection and one-class classification approaches on both known and unknown stego algorithms.