Image steganalysis based on attention augmented convolution

Image steganalysis based on attention augmented convolution

연구 분야: Strategies

학회: Multimedia Tools and Applications

Steganalysis, as the opposite technique to steganography, has been applied to determine whether secret information is embedded in an image. The existing adaptive steganography methods embed secret information in different regions of an image with different probabilities. However, it is difficult for most steganalysis models to make a targeted attention of steganography regions of images, which reduces the detection accuracy of the steganalysis models when detecting the adaptive steganography methods. Aiming to detect the JPEG-based adaptive steganography, this paper proposes a model called Steganalysis Attention Augmented Network (SAANet). The proposed model uses attention augmented convolution instead of traditional convolution, so that the network assigns more learning weights to the steganographic area, and guides the network to better learn features that are beneficial to steganalysis. Thereby improving the learning ability and training effect of the model. As far as we know, we are the first to directly apply attention to the backbone structure of the network model to assist in image steganalysis. And our work improves the accuracy and operating efficiency of the steganalysis model. Experiments show that compared with the current steganalysis models, the model proposed in this paper obtains a competitive detection performance. The current adaptive steganography algorithm achieves a detection accuracy of up to 96.68%, while 95.22% under mismatch conditions, and the testing time is 28 s, indicating that the proposed model has a certain generalization performance and practicability.