1 School of Computer Science and Engineering, Changshu Institute of Technology, No.99, Hushan Road, Changshu 215500, Jiangsu Province, China 2 National Digital Switching System Engineering and Technology Research Center, No.62, Kexue Avenue, Zhengzhou 450001, Henan Province, China
The best traditional steganalysis methods aiming at adaptive steganography are the combination of rich models and ensemble classifier. In this study, a new steganalysis method for JPEG images based on convolutional neural networks is proposed to solve the high dimension problem in steganalysis from another aspect. On the basis of the original rich model, the algorithm adds different sizes of discrete cosine transform (DCT) basis functions to extract different detection features. Different features are combined at the fully connected layer through inputting 2-D feature values to the neural network convolutional layer for predictive classification. Experimental results show that convolutional neural networks as classifiers do not require a large number of training samples, and the final classification performance is better than that of the original ensemble classifier.
