Article Preview
TopIntroduction
Digital image steganography aims to transmit data secretly by embedding secret data into cover image (Zhang, 2016). Nowadays, the plentiful images transmitted over the social network provide convenience for steganography. How to break steganography is becoming a troublesome issue. Steganalysis attempts to reveal the presence of the embedded data. As shown in Figure 1, however, because of the missing detection error of steganalysis, the images judged as “clear” by steganalysis are still possible carried small amount of secret data. Therefore, some slight modifications should be made to prevent the still possible covert communication. On the other hand, the images judged as “stego” by steganalysis are still possible innocent because of the false alarm error. There are many kind of image processing operations frequently used on the images transmitted over social network, such as denoising, recompression, and beautification. So, the false alarm error may be caused by these image processing operations instead of steganography. In this case, it is inappropriate to intercepte all the “stego” images. So, these images can also be processed as same as the “clear” image to stop the quite possible covert communication if the evidence is not enough to declare the guilty of the images. In this way, the combat against steganography is victorious, and meanwhile, the innocent images are transmitted over the social network as usual.
In addition, for early steganographic methods which increase the security performance by decreasing the quantity of embedding changes (Frdrich & Soukal, 2006;Zhang & Wang, 2006; Zhang, Zhang & Wang, 2008), current machine learning based steganalytic methods (Kodovsky, Fridrich & Holub, 2012; Fridrich & Kodovsky, 2012; Holub & Fridrich, 2014; Denemark, Sedighi, Holub, Cogranne & Fridrich, 2014; Song, Liu, Yang, Luo & Zhang, 2015) perform excellent detectability. But for modern steganographic methods (Holub & Fridrich, 2013, Li, Wang, Huang & Li, 2014, Sedighi, Fridrich & Cogranne, 2015, Guo, Ni, Su, Tang & Shi, 2015; Wang, Zhang & Yin, 2016) which improve security performance by minimize the additive distortion between a given cover object and its stego version (Filler, Judas & Fridrich, 2011, Filler & Fridrich, 2010), steganalysis becomes powerless to verdict the presence of secret data especially for the case of small capacity. Recently, adaptive steganalysis (Denemark, Boroumand & Fridrich, J. 2016; Yu, Li, Cheng, Zhang, 2016; Tang, Li, Luo & Huang, 2016) improves the detectability observably. In adaptive steganalysis, different weights are assigned to different cover elements in feature extraction. For the elements with high modifying probabilities, larger weights are assigned since these elements contribute more to steganalysis and vice versa. For small capacity, however, these methods still not perform satisfactory detectability. The detection for small capacity is still a to be resolved problem. In other words, the approach to break steganography is still undiscovered.
Figure 1. The combat against steganography
Actually, to break the covert communication of steganography, steganalysis is not the only choice. For a suspicious image, the possibly existing secret data can be destroyed by modifying the image although it is difficult to judge whether the image is stego or not. In this way, there is no secret data can be transmitted via the modified image. Thus, the threat from steganography is disappeared.