Robust Image Steganography Approach Based on Edge Detection Combined With CNN Algorithm

Image steganography conceals secret data within a cover image to generate a new image (stego image) in a manner that makes the secret data undetectable. The main problem in image steganography is to find the right balance between the maximum amount of secret data that can be transmitted or stored in the cover image and the quality of the stego image to make the secret data un-noticeable to human senses or other detection mechanisms. In this research, a new framework is proposed that integrates the edge detection strategy (using edge detectors) with the deep learning methods, such as a convolutional neural network (CNN), for making secret data embedding and extraction processes efficient. Firstly, the edges in the cover image are identified using a suitable edge detection method (i.e., using the canny or sobel algorithm), and then the secret data is embedded inside the edge-detected cover image using a deep learning approach, and finally, the created stego image is sent to the receiver. On the receiver side, the secret data is extracted from the stego image using the same deep learning model in a reverse manner. In this article, we considered three datasets, such as the Ting ImageNet, Bossbase, and USC-SIPI datasets, to make edge-detected cover images and then consider them to build a deep learning model. We then evaluated the performance of our proposed deep learning model based embedding and extraction approach using various metrics related to imperceptibility, capacity, and robustness. Experimental results show high imperceptibility with PSNR (Peak Signal-to-Noise Ratio) reaching up to 39.85 dB and SSIM (Structural Similarity Index) up to 0.997 when using sobel, and PSNR up to 33.08 dB and SSIM up to 0.995 when using canny. The developed model also provides a higher capacity of 8 bits per pixel (BPP) and have ensured the robustness of the system against common image processing attacks (e.g., JPEG compression, Gaussian noise, salt-and-pepper, speckle, and filtering) which is validated with PSNR values up to 76.2 dB and error rates (Re) ranging between 0.32 and 0.47. Security testing using state-of-the-art steganalyzers (Xu-Net and Ye-Net) shows detection accuracy around 0.45-0.67 only, indicating strong resistance to steganalysis.