A blind steganalysis method based on multi-domain features was proposed for JPEG images to improve the detection accuracy of blind detection scheme for image steganography. The joint probability density matrix was used to capture the correlations of neighboring coefficients in discrete cosine transform (DCT) , and those in discrete wavelet transform (DWT) and spatial domain as the original features. In the same way, the calibrated features were extracted from the calibrated image. The differences between the original and calibrated features were considered to be the exact features used for training classifiers. Experi-ments were done for five kinds of typical JPEG steganography schemes in three different image libraries. Results validated that the detection accuracy of the proposed scheme was greater than 72% when the em-bedded ratio was only 0. 05 bpnc. Compared with some classical blind steganalysis schemes, the proposed method provided higher detection accuracy. Experimental results demonstrate that the feature reduction is an important step in the blind detection framework, which can decrease the correlations of features of high dimensions, improve the detection accuracy and reduce the time of classification greatly.%为了提高图像信息隐藏盲检测的检测率,实现了一种从多个域提取统计特征的JPEG图像通用隐写分析算法.采用联合概率密度来计算离散余弦变换(DCT)系数间的相关性,将其扩展到小波域和空域,提取特征值；并使用同样的方法提取校准图像的特征值.二者的差值组成最终的特征向量训练分类器.对5种典型的JPEG隐写术,使用3种不同图像库进行了一系列的实验,结果表明:在嵌人率只有0.05 bpnc时,该算法检测率大于72％,与目前几种典型的通用隐写分析算法相比,具有更好的检测效果；特征约简能弱化高维特征向量间的相关性,提高分类精确度并大大节约分类时间,是通用隐写分析系统框架中一个重要的步骤.
展开▼
