Research on One-Class JPEG Steganalysis Based on Dimensionality-Reduced Correlation Features

This paper proposed a novel one-class steganalysis method to blindly detect the existence of hidden messages in JPEG images. We used the co-occurrence matrix to capture the correlations among neighboring coefficients in both Discrete Cosine Transform (DCT) domain and Discrete Wavelet Transform (DWT) domain. Then the correlation features were calibrated, their dimensionality was reduced by Locality Preserving Projection (LPP) method and a one-class Support Vector Data Description (SVDD) classifier was trained to make classification. The new method trained only on examples of cover images. The results show that it's able to afford reasonable accuracy to distinguish between cover and stego images. Furthermore, LPP method is much better than Principal Components Analysis (PCA) method for improving the algorithm's clasification accuracy.