<![CDATA[Handcrafted feature based steganalysis remains widely used in resource constrained environments despite rapid progress in deep learning detectors. This study investigates the performance limit of compact handcrafted features for binary classification of cover and stego images in Least Significant Bit (LSB) steganalysis on the Cassavia dataset. Five representative models deep neural network (DNN), one dimensional convolutional neural network (1D CNN), random forest, Light Gradient Boosting Machine (LightGBM), and SMOTE enhanced DNN are trained on 44,000 images using 16 descriptors that combine statistical LSB measures with a reduced subset of Spatial Rich Model (SRM) residual features. All models converge to a narrow accuracy band of 72.58-75.50% with Area Under Curve (AUC) values close to 0.50 and pronounced overfitting in the training–validation curves, indicating that the dominant bottleneck arises from limited feature expressivity rather than model capacity or implementation errors. Feature importance analysis further reveals that only a small subset of descriptors contributes substantially, exposing strong redundancy in the handcrafted feature set. Within this CPU friendly LSB based setting, these results establish a practical performance ceiling that is shared across both classical and deep models, while highlighting LightGBM as an attractive option for embedded steganalysis and motivating future hybrid designs that combine handcrafted statistical priors with learned deep representations.]]>
Copyrights © 2026
