Vector Separability Measurement Based Fast Feature Selection for Detecting Images Information Hiding

Abstract

After careful consideration, we have revised the abstract to “Images are adopted as information hiding carries in recent years, which has brought challenges to multimedia forensics and security. In order to improve the feature selection speed while maintaining the detection accuracy and reducing the feature dimension, this paper proposes a fast feature selection method for detecting stego image with hiding information. First, a feature component separability criterion is proposed based on the inter-distance and inner-distance difference to measure the separability of steganalysis feature components, providing a basis for constructing candidate feature components. Second, the high-dimensional steganalysis feature is extracted from the training set containing the cover images and corresponding stego images. Then the separability value of each feature component is measured based on the above criterion. And candidate feature vectors are constructed based on the separability values of feature components. Third, a feature vector separability measurement, providing a direct basis for the feature component selection to measure the contribution of the candidate feature vector to the classified image classification. The candidate feature vector with the largest value is chosen as the final selection feature. The dimension of the feature selected by this method is exceedingly smaller than the original feature dimension, so the time cost of the subsequent classification training can be significantly reduced. A series of experiments based on the Bossbase image database have shown that: compared with the recent steganalysis-α method and the FSSG method, the proposed method can significantly improve the feature selection speed while maintaining the detection accuracy and reducing the feature dimension.”