Comparison of DCT and Gabor Filters in Residual Extraction of CNN Based JPEG Steganalysis

Abstract

An effective feature selection method to capture the weak stego noise is essential to image steganalysis. In the conventional JPEG steganalysis, Gabor filter and DCT filter are both used for residual extraction. However, there are few comparisons in existing convolutional neural networks (CNNs) based JPEG steganalysis using Gabor filter or DCT filter in the pre-processing stage to extract residuals. In this paper, we compare the performance of DCT filter with Gabor filter in the pre-processing phase of the steganalysis CNN. Firstly, we choose the parameters empirically and theoretically for Gabor filters which are used in CNN. Secondly, we improve the performance by removing the ABS layer in the original XuNet. Finally, the experimental results show that using Gabor filters or DCT filter can achieve comparable performance whenever the parameters of pre-processing filters are fixed or learnable. It’s different from the conventional steganalysis method where Gabor filters have advantages over DCT filters. When the parameters of the pre-processing filters are learnable, both Gabor filter and DCT filter can achieve better performance compared with the condition where the parameters are fixed.