面向隐写分析的图像富模型特征的改进  

Steganalysis Oriented Improved Rich Image Model Features

作　　者：赵宝琴[1] 袁志民[2] ZHAO Bao-qin YUAN Zhi-min(Hebei University of Economics and Business1 , Shijiazhuang 050061, P. R. China Hebei University of Science and Technology2, Shijiazhuang 050018, P. R. China)

机构地区：[1]河北经贸大学信息技术学院,石家庄050061 [2]河北科技大学信息科学与工程学院,石家庄050018

出　　处：《科学技术与工程》2016年第31期56-60,65,共6页Science Technology and Engineering

基　　金：河北省科技计划项目(12210721)资助

摘　　要：针对内容自适应隐写的最佳检测器是经载体图像集与相应隐写图像集训练的集成分类器,训练图像由基于残留噪声的富模型(一族特征)表示。最近研究显示,通过在富模型特征中融入对载体像素的嵌入修改概率,这种内容自适应要素可以提高检测准确度。由于每个残噪样值依赖其周边一整块像素,因此应把对残噪本身而不是对决定残噪的像素的嵌入影响融入富模型特征之中。基于这种认识,提出用残留噪声L1失真的期望值取代像素的嵌入修改率以提高检测准确度。针对当前三种先进的内容自适应隐写算法进行实验,这种新的改进思想得到了实验结果的支持。Presently, the best detectors of content-adaptive steganography are constructed as ensemble classifi- ers trained on sets of cover and stego images. The images are represented with rich models （ a family of features）. Recent research has shown that the detection accuracy can be improved by including adaptive element i. e. the em- bedding change rates in the features. Since each noise residual relies on an entire pixel block, the embedding im- pact on the residual itself rather than on the pixel should be included. According to this observation, the expected value of the residual L1 distortion in the features of rich models in place of the pixel change rates to improve the de- tection accuracy was used. This new idea is supported in experiments for three advanced content-adaptive stegano- graphic algorithms.

关 键 词：隐写分析 富模型 残留噪声 

分 类 号：TP391.4-1[自动化与计算机技术—计算机应用技术]