Cnn-Based Laplacian Residual Network for Image Steganalysis

Abstract

The rapid growth of digital communication has made identifying hidden information in images crucial for preventing misuse and ensuring data security. LR-Net (Laplacian Residual Network) is a CNN-based steganalysis model designed to classify images as either cover or stego. To train and evaluate the proposed model, a dataset is created by generating stego images from cover images using linear interpolation and Tian's Difference Expansion technique. To enhance subtle embedding patterns, a Laplacian high-pass filter is applied before feature extraction. The Laplacian-filtered images are then provided as input to LRA-Net, which learns discriminative features through multiple convolutional layers with batch normalization and ReLU activation. Max pooling reduces the spatial dimensions of the feature maps, while global average pooling followed by fully connected layers improves classification performance and reduces overfitting. The proposed model predicts the probability of an image being either cover or stego, and an optimal threshold is selected to improve classification accuracy. Experimental results demonstrate that the proposed model achieves an accuracy of 85% and reliably distinguishes between cover and stego images, demonstrating its effectiveness for image steganalysis.