A hybrid edge-guided Fourier transform image steganography framework with CNN-Based Detection

Abstract

Image steganography focuses on hiding hidden information in digital images without degrading the quality of images and avoiding any possibility of detecting the hidden message. Traditional edge-based methods adopt predefined algorithms like Sobel and Canny for detecting the regions suitable for embedding, which may not work efficiently if the images have weak and non-clear edges. Besides, traditional spread spectrum techniques require predefined pseudo-noise (PN) sequences as secret keys, which may pose a threat to security and robustness against any geometric transformations. To address these limitations, This paper proposes a secure image steganography framework that integrates Hybrid Edge-Guided Fourier-Domain Steganography Framework with the Convolutional Neural Network (CNN) based learned detector for securing the process of image steganography. A content adaptive attention mechanism will be adopted to detect the best regions of embedding in the spatial domain, whereas PN classes embedding will be done in the Fourier domain. The CNN based encoder-decoder network will be used to hide and recover the data, and a trained CNN classifier will allow a blind detection. Efficiency and security of this approach will be measured by calculating PSNR, SSIM, MSE, BPP and Re using Xu-Net and Ye-Net steganalysis models.