Wavelet-Based Fusion for Image Steganography Using Deep Convolutional Neural Networks

Abstract

Steganography has long served as a powerful tool for covert communication, particularly through image-based techniques that embed secret information within innocuous cover images. With the increasing adoption of deep learning, researchers have sought more secure and efficient methods for image steganography. This study builds upon and extends the DeepWaveletFusion approach by integrating convolutional neural networks (CNNs) with the discrete wavelet transform (DWT) to enhance both embedding and recovery performance. The proposed method, DeepWaveletFusionToo, is a lightweight architecture that employs a custom-built DWT image dataset and leverages the mean squared error (MSE) loss function during training, significantly reducing model complexity and computational cost. Experimental results demonstrate that DeepWaveletFusionToo achieves improved imperceptibility compared to its predecessor and delivers competitive recovery accuracy over existing deep learning-based steganographic techniques, establishing its simplicity and effectiveness.