Research map / Image generation / Image restoration
Purified and Unified Steganographic 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 PUSNet, a sparse image denoising network that hides secret image encoder and decoder networks in its sparse weights, triggered by keys via key-seeded weight filling.
Problem
Transmitting large DNN-based steganographic networks (over 100MB) covertly to senders and receivers, without side information and while concealing both encoding and decoding networks inside an ordinary-looking DNN model.
Approach
Formulates DNN model steganography as sparse weight filling: a purified sparse denoising network keeps its masked weights, and a key seeds random Xavier weights that fill the masked positions to produce either a secret encoder or decoder; sparse masks come from magnitude-based pruning at initialization, and only shared sparse weights are trained.
Key results
- PUSNet-E achieves stego-image PSNR of 38.15 dB on DIV2K, over 9.73 dB above Baluja and HiDDeN, while HiNet reaches 44.86 dB.
- StegExpose AUC of 0.58 against stego images, close to random guessing; SiaStegNet needs at least 100 cover/stego pairs to detect.
- PUSNet-P denoising PSNR on COCO is 32.95 dB versus 33.68 dB for the pure denoiser PUSNet-C, a reduction under 0.8 dB; PUSNet-D recovery is 26.96 dB PSNR versus 28.52 dB for Li et al. [18].
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PUSNet-E instead of HiNet for image embedding | PUSNet · DIV2K, COCO, and ImageNet cover/stego-image pairs | PSNR (dB): 44.86 vs 38.15; 46.47 vs 39.09; 46.88 vs 38.94 | Hurt |
| Use PUSNet-D instead of UDH for secret recovery | PUSNet · DIV2K, COCO, and ImageNet secret/recovered image pairs | PSNR (dB): 30.52 vs 26.88; 30.52 vs 26.96; 29.63 vs 26.28 | Hurt |
| Use the sparse PUSNet-P instead of the denoising PUSNet-C | PUSNet image denoising · DIV2K, COCO, and ImageNet | PSNR (dB): 33.03 vs 32.25; 33.68 vs 32.95; 33.66 vs 32.96 | Hurt |
| Trigger the secret encoder with a randomly guessed key | PUSNet-P · DIV2K, COCO, and ImageNet random-key secret embedding | PSNR (dB): 8.81±1.74; 8.01±2.07; 7.74±1.89 | Hurt |
| Trigger the secret decoder with a randomly guessed key | PUSNet-P · DIV2K, COCO, and ImageNet random-key secret recovery | PSNR (dB): 6.52±0.78; 6.25±0.97; 6.73±0.83 | Hurt |
| Use PUSNet-D instead of Li et al. [18] | Steganographic network hiding · COCO secret recovery | PSNR (dB): 28.52 vs 26.96 | Hurt |
| Use PUSNet instead of Li et al. [18] for hiding the steganographic network | Steganographic network hiding · Image denoising | PSNR performance reduction (dB): 1.24 vs 0.73 | Helped |
Surprising findings
- Randomly guessed keys produce poor stego and recovered images (PSNR under 9 dB), so random filling attacks largely fail.
- Under the weight-distribution EMD test, PUSNet-P is closer to PUSNet-C1 than PUSNet-C1 is to PUSNet-C2, so the weight distribution gives no reliable detection signal.
Limitations
- Secret recovery quality (PSNR about 26.96 dB) is lower than the prior Li et al. method (28.52 dB) and HiNet.
- Stego images have lower PSNR than HiNet (38.15 dB vs 44.86 dB on DIV2K).
- Detection tests for DNN model steganalysis are empirical heuristics; no formal security guarantee is given.
Experimental setup
Datasets: DIV2K
Benchmarks: DIV2K COCO ImageNet
Metrics: PSNR SSIM APD RMSE AUC EMD
Baselines: HiDDeN Baluja UDH HiNet Li et al. model steganography
Builds on: HiNet HiDDeN UDH Baluja steganography magnitude-based pruning at initialization
Closest papers by meaning
- Steganography of Steganographic Networks (AAAI 2023 · similarity 0.86)
- Efficient and Separate Authentication Image Steganography Network (ICML 2025 · similarity 0.84)
- Hiding Images in Diffusion Models by Editing Learned Score Functions (CVPR 2025 · similarity 0.82)
- Robust Message Embedding via Attention Flow-Based Steganography (CVPR 2025 · similarity 0.82)
- StegFormer: Rebuilding the Glory of Autoencoder-Based Steganography (AAAI 2024 · similarity 0.81)
- SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images (ICML 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.