DCT-Based Steganalysis for Steganographic Algorithm and Distortion-Level Classification Using Deep Learning

This study proposes a methodology based on Discrete Cosine Transform (DCT) coefficients and deep learning to identify steganographic algorithms (UERD, J-UNIWARD, and JMiPOD) and to classify distortion levels in images from the ALASKA2 dataset. For each image, 258,048 AC coefficients were extracted, and 9999 were selected using Mutual Information (MI), allocated equally across three frequency bands. Two representations were compared under identical architectures: the MI-selected coefficients and full spatial maps of the DCT coefficients. Four architectures (CNN-1D, CNN 2D, CNN 2D + CBAM, and Extended Deep CNN) were trained with three random seeds. Spatial maps outperformed the MI-selected coefficients for every architecture, by 11.7 percentage points in algorithm identification (CNN 2D) and by 1.0 to 2.1 points in distortion-level classification. The best results were obtained by 2D convolutional networks on spatial maps: 59.2 ± 0.2% accuracy for algorithm identification (CNN 2D + CBAM, 0.8 points above the same network without attention) and 66.5 ± 0.5% for distortion levels (CNN 2D). A network with about eleven times more parameters did not improve the results. Distinguishing among the considered algorithms and distortion levels remains challenging.