文摘
Despite the state-of-the-art steganalysis can detect highly undetectable steganography, it is too unreliable to implement in the real world due to its false alarm rate. In pooled steganalysis scenario, multiple objects are intercepted and a reliable collective decision is required. To control the reliability, the confidence intervals of the detectors’ false rates are estimated as a parameter and hypothesis testing technology is used to determine the threshold of stego rates. In view of the fact that the false rate is vulnerable to some image properties (e.g. image size, and texture complexity), we propose a novel fine-grained scheme where test sets are divided by its texture measure in both parameter estimation and hypothesis testing processes. The demonstration on public image sets shows the proposed scheme achieves higher reliability in most cases. It confirms that the priori knowledge of image properties is conductive to a accurate threshold and reliable decision.