Ensemble of CNN and rich model for steganalysis
Kai Liu, Jianhua Yang, Xiangui Kang · 2017
Recent studies have indicated that well-designed convolutional neural network (CNN) has achieved comparable performance to the spatial rich models with ensemble classifier (SRM-EC) in digital image steganalysis. In this paper, we discuss the difference and correlation between a CNN model and a SRM-EC model, and explore the classification error rate varying with texture complexity of an image for both models. Then we propose an ensemble method to combine CNN with SRM-EC by averaging their output classification probability. Compared with the state-of-the-art performance of spatial steganalysis achieved by maxSRMdZ, which is the latest variant of SRM-EC, experimental result shows that the proposed ensemble method furtherly improves the accuracy by nearly 2% in detecting S-UNIWARD and WOW on BOSSbase.