Steganography is the task of hiding information in some media normally images. Steganalysis is the process of discriminating such instances and clean ones. In recent years, steganalysis has tended to use deep learning for feature extraction and classification. Convolutional Neural Networks (CNN) have improved the steganalysis performance but at the cost of computational complexity and memory space due to huge amount of training data. In this paper, a new framework is proposed to reduce the learning cost by a divide and conquer strategy. In the first phase, data is divided into disjoint clusters by use of k-means. Each cluster is then fed to a separate CNN to be customized on a specific region of data space. In the final phase, the networks are merged leveraging a fast alternate-weighting process. The proposed weighting can, to some extent, compensate for reducing the size of training data per model. The experimental results show that the proposed scalable framework reduces memory and time complexity with preserving accuracy.
展开▼
