Research map / Image generation / Autoregressive and latent generation
Efficient and Separate Authentication Image Steganography Network
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 AIS, a two-stage invertible image steganography network that embeds lock-key authentication via a separate invertible authentication network and hides fused information in parallel with a single hiding network.
Problem
Multi-image steganography reveals all hidden secret images without recipient authentication; adding authentication information degrades stego/secret quality and increases model size for large-capacity hiding.
Approach
Separate the authentication step from hiding: an Invertible Authentication Network fuses locks with secret images and extracts primary information, then a single Invertible Hiding Network hides the fused information in parallel, reducing cover space usage and model size.
Key results
- On DIV2K hiding 5 secret images, AIS reaches 36.149 dB PSNR cover-stego versus 26.735 dB for ISN and uses 6.15M parameters versus 55.62M for IIS.
- With 3 secret images on DIV2K, AIS achieves 34.721 dB stego PSNR and 34.761 dB revealed-secret PSNR, outperforming ISN, DeepMIH and IIS.
- Detection accuracy against SRNet, ZhuNet and LWENet drops to 52.80%, 61.40% and 51.75%, the lowest among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AIS instead of the strongest baseline | Image steganography with 5 secret images · DIV2K, cover-stego | PSNR: 29.477 -> 36.149 | Helped |
| Use AIS instead of the strongest baseline | Image steganography with 5 secret images · DIV2K, secret-reveal | PSNR: 28.676 -> 30.765 | Helped |
| Use AIS instead of the strongest baseline | Image steganography with 5 secret images · ImageNet, cover-stego | SSIM: 0.792 -> 0.841 | Helped |
| Use AIS instead of the strongest baseline | Image steganography with 5 secret images · ImageNet, secret-reveal | SSIM: 0.790 -> 0.833 | Helped |
| Extract 3 channels rather than 12 from secret images (ablation) | AIS · DIV2K, cover-stego | PSNR: 33.488 -> 36.149 | Helped |
| Extract 3 channels rather than 9 from secret images (ablation) | AIS · DIV2K, cover-stego | PSNR: 30.568 -> 36.149 | Helped |
| Extract 3 channels rather than 6 from secret images (ablation) | AIS · DIV2K, cover-stego | PSNR: 29.472 -> 36.149 | Helped |
| Extract 3 channels rather than 12 from secret images (ablation) | AIS · DIV2K, secret-reveal | PSNR: 27.846 -> 30.765 | Helped |
| Extract 3 channels rather than 9 from secret images (ablation) | AIS · DIV2K, secret-reveal | PSNR: 27.751 -> 30.765 | Helped |
| Extract 3 channels rather than 6 from secret images (ablation) | AIS · DIV2K, secret-reveal | PSNR: 26.767 -> 30.765 | Helped |
| Use two-stage distribution adaptation rather than one-stage authentication (ablation) | AIS · DIV2K | PSNR of revealed secret images: +2.086 db | Helped |
| Use two-stage distribution adaptation rather than one-stage authentication (ablation) | AIS · DIV2K | information entropy of hidden information: 1.85 | Helped |
Surprising findings
- Adding authentication information in a one-stage design measurably lowers stego and revealed image quality and increases parameters (16.3M to 19.9M in the authentication-free vs authentication-based comparison).
Limitations
- Single-network parallel hiding relies on separate authentication; quality on ImageNet for some configurations trails baselines (e.g., LPIPS for revealed secrets at N=2)
- Evaluated only on DIV2K and ImageNet with fixed image sizes and up to five secret images
Experimental setup
Datasets: DIV2K ImageNet
Benchmarks: DIV2K ImageNet
Metrics: PSNR SSIM LPIPS JS divergence FLOPs inference time steganalysis detection accuracy
Baselines: ISN DeepMIH IIS
Builds on: IIS HiNet DeepMIH ISN conditional normalizing flows
Closest papers by meaning
- Large-capacity and Receiver Authenticable Generative Image Steganography (ICML 2026 · similarity 0.88)
- SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images (ICML 2025 · similarity 0.84)
- StegFormer: Rebuilding the Glory of Autoencoder-Based Steganography (AAAI 2024 · similarity 0.84)
- Purified and Unified Steganographic Network (CVPR 2024 · similarity 0.84)
- StegaStyleGAN: Towards Generic and Practical Generative Image Steganography (AAAI 2024 · similarity 0.81)
- Provably Secure Image Robust Steganography via Cross-modal Error Correction (AAAI 2025 · similarity 0.81)
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.