[Paper Review] Color Image steganography using Deep convolutional Autoencoders based on ResNet architecture

Seyed Hesam Odin Hashemi, Mohammad‐Hassan Majidi+1 more|arXiv (Cornell University)|Nov 17, 2022

Advanced Steganography and Watermarking Techniques24 references4 citations

TL;DR

This paper proposes a deep learning-based color image steganography method using a ResNet-inspired convolutional autoencoder for high-capacity, secure, and imperceptible data hiding. The model uses a shared architecture for both embedding and extraction, achieving PSNR >40 dB and SSIM >0.98 on COCO and CelebA datasets, with a hiding capacity of 8 bits per pixel for full-color images.

ABSTRACT

In this paper, a deep learning color image steganography scheme combining convolutional autoencoders and ResNet architecture is proposed. Traditional steganography methods suffer from some critical defects such as low capacity, security, and robustness. In recent decades, image hiding and image extraction were realized by autoencoder convolutional neural networks to solve the aforementioned challenges. The contribution of this paper is introducing a new scheme for color image steganography inspired by ResNet architecture. The reverse ResNet architecture is utilized to extract the secret image from the stego image. In the proposed method, all images are passed through the prepossess model which is a convolutional deep neural network with the aim of feature extraction. Then, the operational model generates stego and extracted images. In fact, the operational model is an autoencoder based on ResNet structure that produces an image from feature maps. The advantage of proposed structure is identity of models in embedding and extraction phases. The performance of the proposed method is studied using COCO and CelebA datasets. For quantitative comparisons with previous related works, peak signal-to-noise ratio (PSNR), the structural similarity index (SSIM) and hiding capacity are evaluated. The experimental results verify that the proposed scheme performs better than traditional and pervious deep steganography methods. The PSNR and SSIM are more than 40 dB and 0.98, respectively that implies high imperceptibility of the proposed method. Also, this method can hide a color image of the same size in another color image, which can be inferred that the relative capacity of the proposed method is 8 bits per pixel.

Motivation & Objective

• To address the limitations of traditional steganography methods, including low capacity, poor security, and limited robustness.
• To develop a deep learning framework that enables high-fidelity embedding and extraction of secret color images within cover images.
• To leverage the residual learning structure of ResNet to improve feature representation and reconstruction quality in steganographic systems.
• To achieve identity in the embedding and extraction models for simplified deployment and enhanced consistency.
• To evaluate the method on standard datasets (COCO and CelebA) using metrics like PSNR, SSIM, and hiding capacity.

Proposed method

• A pre-processing model based on a convolutional deep neural network is used to extract features from both cover and secret images.
• An operational model, structured as a ResNet-based autoencoder, generates stego images from the extracted features.
• The autoencoder employs skip connections inspired by ResNet to stabilize training and improve gradient flow during end-to-end learning.
• The same model architecture is used for both embedding and extraction, ensuring symmetry and simplifying implementation.
• The method operates in an end-to-end fashion, learning to map secret images to stego images while preserving visual quality.
• The network is trained to minimize reconstruction loss between the original secret image and the extracted image, using perceptual and structural similarity metrics.

Experimental results

Research questions

• RQ1 Can a ResNet-based deep autoencoder achieve high imperceptibility in color image steganography while maintaining high hiding capacity?
• RQ2 Does the use of a shared model architecture for embedding and extraction improve system consistency and simplify deployment?
• RQ3 How does the proposed method compare to traditional and prior deep learning-based steganography methods in terms of PSNR and SSIM?
• RQ4 What is the maximum hiding capacity achievable using this architecture for full-color images?
• RQ5 Can the method effectively preserve structural and perceptual quality when embedding a secret image of the same size as the cover image?

Key findings

• The proposed method achieves a peak signal-to-noise ratio (PSNR) of over 40 dB, indicating high visual fidelity and imperceptibility of the stego image.
• The structural similarity index (SSIM) exceeds 0.98, confirming that the stego image preserves structural details of the original cover image.
• The hiding capacity reaches 8 bits per pixel, enabling the embedding of a full-color secret image of the same size as the cover image.
• The method demonstrates superior performance compared to traditional and prior deep learning-based steganography techniques in both quantitative and qualitative evaluations.
• The symmetric architecture used in both embedding and extraction phases ensures consistent and reliable reconstruction of the secret image.