Source-linked AI summary
End-to-end Trained CNN Encode-Decoder Networks for Image Steganography
Atique ur Rehman, Rafia Rahim, M Shahroz Nadeem, Sibt ul Hussain
TL;DR
Existing steganography methods mainly hide binary payloads using manually crafted features, limiting the paper’s target problem to lower-capacity, distortion-sensitive embedding. The paper introduces jointly trained CNN encoder-decoder networks that embed an image into a cover image and recover it. Across several datasets, it reports high payload capacity with high PSNR and SSIM, including 33.3% payload at 32.92 dB and 36.58 dB PSNR on ImageNet.
Problem
Existing methods mainly hide binary payloads using manually crafted image features, whereas image-in-image payload embedding is the paper’s target problem.
Method
A CNN encoder-decoder is trained end-to-end with a joint loss to create a hybrid cover image and recover the embedded guest image.
Results
33.3% payload capacity was maintained on ImageNet with 32.92 dB PSNR for the encoder output and 36.58 dB for the decoder output, both at SSIM = 0.96.
Takeaways & Limitations
The method directly supports image payloads and is evaluated across MNIST, CIFAR10, PASCAL-VOC12, ImageNet, and LFW.
Abstract
from arXiv · showhide
All the existing image steganography methods use manually crafted features to hide binary payloads into cover images. This leads to small payload capacity and image distortion. Here we propose a convolutional neural network based encoder-decoder architecture for embedding of images as payload. To this end, we make following three major contributions: (i) we propose a deep learning based generic encoder-decoder architecture for image steganography; (ii) we introduce a new loss function that ensures joint end-to-end training of encoder-decoder networks; (iii) we perform extensive empirical evaluation of proposed architecture on a range of challenging publicly available datasets (MNIST, CIFAR10, PASCAL-VOC12, ImageNet, LFW) and report state-of-the-art payload capacity at high PSNR and SSIM values.
1. INTRODUCTION
The paper targets image-in-image steganography, extending beyond binary payloads and manually selected image features. It proposes an automatic CNN approach that embeds one image into another and evaluates it across diverse datasets with high payload capacity and perceptual scores.
- Prior work: Existing image steganography primarily hides binary or text payloads by exploiting manually selected edges, textures, noisy regions, or other low-level features.These approaches aim to maximize embedded information while limiting visible distortion.
- Problem and novelty: The paper formulates hiding a real-world image payload inside a real-world cover image as a generic steganography problem.Text can also be converted into a bitmap and used as an image payload.
- Proposed approach: The proposed method automatically learns features from cover and payload images to merge their information through a deep learning network.The authors describe the method as novel and completely automatic.
- Headline results: 33% payload capacity is reported at average PSNR values of 32.9 dB for the cover and 36.6 dB for the recovered payload, with SSIM = 0.96 for both.The reported average payload is 8 bpp.
2. METHODOLOGY
The method jointly trains CNN encoder and decoder networks to create a visually similar hybrid image from host and guest images and then recover the guest. Its loss combines host-image fidelity, guest-image reconstruction, and weight regularization, while training is end-to-end and testing uses the networks separately.
- Network roles: CNN feature hierarchies let the encoder identify cover-image features for hiding payload details, while the decoder separates those details from the hybrid.The encoder and decoder are trained as a pair.
- Network roles: The encoder accepts host and guest images and produces a hybrid image that remains visually similar to the host while containing guest content.The decoder receives the hybrid image and reconstructs the guest image.
- Joint objective: The joint loss combines hybrid-image fidelity, recovered-guest fidelity, and encoder and decoder weight penalties.α and β control the encoder and decoder terms, while We and Wd denote learned network weights.
- Encoder architecture: The encoder uses parallel guest and host branches that progressively merge guest features into host features.The branches are combined through alternating-layer feature-map concatenation, with merging continuing to depth k = 7 in the reported design.
- Decoder architecture: The decoder applies convolutional and ReLU layers to the encoder output to recover the concealed guest representation.The final layer omits ReLU according to the architecture description.
- Training and testing: Training uses the joint loss end-to-end, whereas testing uses the encoder and decoder in a disjoint manner.The paper states that alternative design choices were tested, with this architecture selected from initial experiments.
3. EXPERIMENTS AND RESULTS
The experiments evaluate the same CNN steganography approach across diverse datasets, measuring perceptual quality and embedding capacity with PSNR, SSIM, and bpp. Results show image payloads can be hidden and recovered at substantial capacity across simple and complex datasets, including unseen datasets.
- Evaluation metrics: PSNR, SSIM, and bits per pixel (bpp) measure perceptual image quality and embedding capacity.
- CIFAR10 results: 33.3% payload capacity (8 bpp) was achieved on CIFAR10 with average PSNR values of 30.9 dB for encoder output and 29.9 dB for decoder output.Both cover and payload images were 32 × 32, with RGB covers and single-channel payloads.
- ImageNet results: 33.3% payload capacity (8 bpp) was achieved on ImageNet, with final PSNR values of 32.92 dB and 36.58 dB and SSIM = 0.96 for encoder and decoder outputs, respectively.The final ImageNet model was trained for 150 epochs and maintained similar payload capacity.
- Cross-dataset generalization: The ImageNet-trained algorithm achieved high payload capacity at high PSNR and SSIM values on 1,000 unseen PASCAL-VOC12 and LFW images.Figure 2 additionally shows qualitative concealment and recovery results for LFW, PASCAL-VOC12, and ImageNet.
4. CONCLUSIONS
The paper presents a CNN encoder-decoder architecture for hiding image payloads directly inside cover images. Extensive experiments across wild-image datasets report strong payload capacity and image recovery.
- The proposed method directly embeds an image payload into a cover image and uses encoder-decoder networks to recover it.
- The authors report extensive experiments showing strong payload capacity across a wide range of wild-image datasets.