Deep Learning-Based Multi-Class Detection of LSB Steganography in Digital Images

Abstract

Identifying hidden payloads in images has become increasingly critical as steganography continues to challenge traditional security measures. This paper introduces a deep learning framework for both the detection (binary classification) and fine-grained classification (multi-class) of steganographic payloads embedded using Least Significant Bit (LSB) techniques. The proposed system distinguishes between benign images and stego images containing five different payload types: HTML, JavaScript, PowerShell, URLs, and Ethereum-related data. To achieve this, we systematically evaluate various architectures, including a custom Convolutional Neural Network (CNN), hybrid CNN-GRU and CNN-LSTM models, and a Vision Transformer (ViT) at different input resolutions using 5-fold cross-validation. Our experiments reveal a critical finding: image resizing significantly degrades detection performance, as subtle LSB artifacts are often corrupted. While our custom CNN model achieved the highest mean cross-validation accuracy (0.9702), the hybrid CNN-GRU model demonstrated superior generalization on the held-out test set and external dataset, achieving a multi-class accuracy of 0.98 on the test set and 0.97 on the external. This result highlights the advantage of combining the CNN’s spatial feature extraction with the GRU’s ability to model sequential dependencies for robust payload identification on unseen data.