Research map / Image generation / Image restoration
StegFormer: Rebuilding the Glory of Autoencoder-Based Steganography
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes StegFormer, a U-shaped transformer autoencoder for large-capacity image hiding, plus a normalizing training strategy (restrict loss and gradient truncation) and a cascading method for multi-image hiding.
Problem
Large-capacity image hiding (concealing one or more secret images in a same-resolution cover image) has been dominated by INN-based models, which are slow to train and architecturally rigid; existing models also ignore that stego pixel values fall outside [0,255] in real-world transmission, causing
Approach
An autoencoder with Swin-style channel-adaptive transformer blocks in a U-Net layout, a global-attention bottleneck with conditional positional encoding, a restrict loss plus gradient truncation to keep stego values in range, and capacity expansion by cascading secret images.
Key results
- Under ideal conditions on DIV2K, StegFormer reaches 56.30 dB cover/stego PSNR and 55.45 dB secret/recovery PSNR, about 5.5 dB and 3.1 dB above the second-best method.
- Under realistic conditions (stego truncated to [0,1]), normalizing training raises StegFormer's secret/recovery PSNR on DIV2K from 37.89 to 53.43 dB.
- For 2-image hiding on COCO, StegFormer's average recovery PSNR is 40.67 dB versus DeepMIH's 37.14 dB (+3.5 dB); 4-image hiding reaches 33.66 dB average recovery PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use StegFormer instead of HiNet | Single-image hiding · DIV2K, cover/stego image pair | PSNR: 50.79 -> 56.30 | Helped |
| Use StegFormer instead of DAH-Net | Single-image hiding · COCO, cover/stego image pair | PSNR: 46.15 -> 48.77 | Helped |
| Use StegFormer instead of HiNet | Single-image hiding · ImageNet, cover/stego image pair | PSNR: 46.00 -> 48.79 | Helped |
| Use StegFormer instead of HiNet | Single-image hiding · DIV2K, secret/recovery image pair | PSNR: 52.32 -> 55.45 | Helped |
| Use StegFormer instead of DAH-Net | Single-image hiding · COCO, secret/recovery image pair | PSNR: 47.46 -> 49.21 | Helped |
| Use StegFormer instead of DAH-Net | Single-image hiding · ImageNet, secret/recovery image pair | PSNR: 47.35 -> 49.18 | Helped |
| Add normalizing training (restrict loss and gradient truncation) (ablation) | StegFormer · DIV2K, realistic conditions, secret/recovery image pair | PSNR: 37.89 -> 53.43 | Helped |
| Add normalizing training (restrict loss and gradient truncation) (ablation) | StegFormer · DIV2K, realistic conditions, cover/stego image pair | PSNR: 56.30 -> 55.98 | Hurt |
| Use restrict loss and gradient truncation together rather than neither (ablation) | StegFormer · Realistic conditions, secret/recovery image pair | PSNR: 42.26 -> 53.43 | Helped |
| Use restrict loss and gradient truncation together rather than restrict loss alone (ablation) | StegFormer · Realistic conditions, cover/stego image pair | PSNR: 53.56 -> 55.98 | Helped |
| Use restrict loss and gradient truncation together rather than restrict loss alone (ablation) | StegFormer · Realistic conditions, secret/recovery image pair | PSNR: 52.03 -> 53.43 | Helped |
| Add Channel Adapter (ablation) | StegFormer · Cover/stego image pair | PSNR: 53.94 -> 55.98 | Helped |
Surprising findings
- Naively adding a restrict loss without gradient truncation hurts recovery quality (secret PSNR 42.26 dB vs 53.43 dB with both), showing the two components must be used together.
Limitations
- Root cause of real-world failure of current models is left unexplored by the authors.
- Multi-image capacity gains are modest for 4 images (average recovery PSNR 33.66 dB, +0.6 over DeepMIH).
Experimental setup
Datasets: DIV2K COCO ImageNet
Benchmarks: DIV2K COCO ImageNet
Metrics: PSNR SSIM MAE MSE RMSE
Baselines: HiDDeN UDH HiNet DeepMIH DAH-Net
Builds on: BalujaNet HiNet Uformer Swin Transformer CPE
Closest papers by meaning
- SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images (ICML 2025 · similarity 0.87)
- StegaStyleGAN: Towards Generic and Practical Generative Image Steganography (AAAI 2024 · similarity 0.85)
- Efficient and Separate Authentication Image Steganography Network (ICML 2025 · similarity 0.84)
- Provably Secure Image Robust Steganography via Cross-modal Error Correction (AAAI 2025 · similarity 0.84)
- Robust Message Embedding via Attention Flow-Based Steganography (CVPR 2025 · similarity 0.84)
- Neural Cover Selection for Image Steganography (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.