KL-sense secure image steganography

In this paper, we propose a computationally–efficient data hiding method which achieves Cachin's security criterion: zero Kullback–Liebler (KL) divergence. To preserve statistical properties of the cover medium, we swap pixels rather than modify them to hide information. We theoretically analyse the security of the proposed method from various perspectives. Upper bounds of the KL divergence of second order statistics; The relationship between distortions in the DCT domain and embedding positions in the spatial domain; The upper bound on the conditional entropy in the DCT domain. We then subject our proposed stego method to several practical steganalysis algorithms: Histogram based attacks; A higher–order statistics based universal steganalysis algorithm; A new learning based steganalysis that specifically for this hiding algorithm. Experimental results show that our data hiding method can prevent these statistical detection methods, when the embedding rate is less than or equal to 10%.