Unsupervised universal steganalysis combining image retrieval and outlier detection
Chen Guang Xu, Tao Zhang, Xiaodan Hou · 2016 IEEE Information Technology, Networking, Electronic and Automation Control Conference · 2016
Traditional steganalysis framework based on the binary classification is well developed. However, the unknown embedding algorithm and unknown cover source inevitably create so-called model mismatch problem, which leads to the significantly decrease of detection performance in practical applications. This study proposes a new unsupervised universal steganalysis framework that combines image retrieval and outlier detection to address the model mismatch problem. First, the statistical properties of the given test image are initially based to retrieve similar images from a massive cover image dataset as an aided image set. Second, outlier detection is conducted for the test image set, which is composed of the given test image and its aided image set to verify if the given test image is embedded. Experimental results show that the proposed framework can effectively achieve unsupervised universal steganalysis on JPEG heterogeneous image databases and avoid model mismatch.