Performance of Statistical Stegdetectors in Case of Small Number of Stego Images in Training Set

Today special interest is taken to early detection and counteraction to hidden (steganographic) communication between intruders. Such communication based on message embedding into files, for instance digital images, that are processed and transmitted in a communication system. The wide range of steganalysis methods for revealing of formed stego images was developed. Proposed stegdetectors allows achieving high detection accuracy (more than 95%) for most steganographic methods when steganalyst has access to the embedding algorithm. In case of limited a priori information about used steganographic method, when steganalyst can use only a small amount of stego images during stegdetector tuning, detection accuracy may decreases drastically. The paper is devoted to the performance analysis of state-of-the-art stegdetector based on maxSRMd2 statistical model of cover image by limited quantity of available stego images. The case of adaptive embedding of stegodata in cover image according to modern HUGO and S-UNIWARD method is considered. The obtained results indicate a strong dependence of stegdetector performance on number of cover-stego images pairs at detector setup stage, which puts forward additional requirements for training set during detectors tuning.