Machine learning Model for Digital Image Steganalysis

Abstract

Digital image steganalysis plays a vital role in detecting hidden information embedded within images, yet identifying advanced steganographic techniques remains a significant challenge. This paper presents a hybrid feature extraction framework that combines deep features from Residual Network (ResNet)-50 with handcrafted statistical, texture, and transform-domain features to enhance detection performance. Using the ALASKA2 dataset, images are preprocessed and augmented before extracting Subtractive Pixel Adjacency Matrix (SPAM), Local Binary Pattern (LBP), co-occurrence, and wavelet residual features, which are fused with deep representations.  Principal Component Analysis-SHapley Additive exPlanations (PCA)-(SHAP) -based feature selection approach reduces dimensionality while improving interpretability by retaining the most informative features. The selected features are classified using an ensemble of eXtreme Gradient Boosting (XGBoost) and Deep Convolutional Neural Network (DCNN) models. Experimental results demonstrate an accurate score of 98.57%, outperforming recent state-of-the-art methods and highlighting the effectiveness of the proposed model for robust digital image steganalysis.