Deep Feature Extraction Techniques For Machine Learning-Based Digital Image Steganalysis

Abstract

Digital images are increasingly being used to improve communication, disseminate information and support e-commerce. This leads to the increasing demand for effective approaches for detecting hidden information via steganography. We present a machine-learning steganalysis framework, which employs deep feature extraction. It combines deep learning, statistical representation of functions, dimensionality reduction, interpretable feature assessment and ensemble classification. To do this, the suggested approach employs a set of cover and stego photos from the ALASKA2 dataset. Image Preprocessing Image pre-processing includes normalisation, scaling, noise factors and augmentation, which are used to enhance the learning process. In addition, enhanced deep image features extracted using ResNet-50 are combined with wavelet-based residual attributes and statistical features such co-occurrence matrices, Local Binary Patterns (LBP), and Subtractive Pixel Adjacency Matrix (SPAM). Thus, the combined feature representation reduces the feature redundancy and dimension using Principal Component Analysis (PCA). We next use SHAP (SHapley Additive exPlanations) to discover and rank the important feature components correlating with the category, for subsequent classification interpretability. The best feature selection is evaluated using the ensemble classification model of DCNN and XGBoost. The experimental results show that the proposed Ensemble + PCA + ResNet-50 framework achieves high performance of 98.57% accuracy, 99.48% precision, 98.46% recall and 98.97 F1-score as compared to some models used CNN based architectures (Resnet50) and deep learning methods on the dataset provided in details in this paper. Moreover, the proposed approach for extraction and classification is compared with the state of the art approaches for steganalysis to show its effectiveness. This shows that the combination of deep and handcrafted features with PCA optimisation, subjective explainable AI (SHAP) interpretability tool and ensemble methods may build a strong and consistent framework of digital image steganalysis.