| Abstrakt: |
The author analyzes the problem of detecting the adaptive steganography by the J-UNIWARD method by steganoanalytical systems based on machine learning. As determined by the comparative analysis of the accuracy, the statistical models of constructing characteristic vectors that are calculated in the spatial domain, such as GFR, PHARM, and DCTR, are most sensitive to J-UNIWARD. Here, two following ways to improve the accuracy of steganoanalysis based on these models are proposed: via the analysis of the most probable embedding locations and via the balanced vote on the three models. Significant degradation of the accuracy of steganoanalysis without preliminary classification of images according to their parameters is demonstrated. The obtained results can be used to generate efficient steganoanalysis systems for JPEG images. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink