Research map / Image generation / Autoregressive and latent generation
Provably Secure Image Robust Steganography via Cross-modal Error Correction
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 CMSteg, a provably secure, robust image steganography method that embeds messages into VQ tokens sampled from the autoregressive LlamaGen model, with discrete token optimization and cross-modal error correction via InstructBLIP-generated stego text.
Problem
Provably secure image steganography with existing generators yields low-quality, low-resolution images and loses synchronization under lossy social-network processing such as JPEG compression.
Approach
Embed messages via distribution-preserving (Discop-style) sampling in an autoregressive VQ-token image generator, recover tokens at the receiver by re-encoding plus gradient-based optimization in continuous latent space, and send compressed token-error corrections hidden in a descriptive text generated by a vision-language model using provably secure linguistic steganography.
Key results
- Without noise, the full pipeline (M1+M2+M3) reaches 99.98% token recovery and 4285 bits effective capacity versus 91.89% and 106 bits for M1 alone.
- Under JPEG QF 75, token recovery rises from 78.65% (M1) to 98.42% (M1,2,3), with capacity 3293 bits.
- Steganalyzers ConvNet, LWENet, and SRNet give detection error rates of 0.5014, 0.5043, and 0.4980, near chance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add discrete token optimization (M2) (ablation) | CMSteg image steganography · Lossless channel | Cap: 106 -> 3803 | Helped |
| add cross-modal error correction (M3) (ablation) | CMSteg image steganography · Lossless channel | Cap: 3803 -> 4285 | Helped |
| add cross-modal error correction (M3) (ablation) | CMSteg image steganography · JPEG QF 95 | Cap: 2793 -> 3936 | Helped |
| add cross-modal error correction (M3) (ablation) | CMSteg image steganography · JPEG QF 85 | Cap: 2551 -> 3861 | Helped |
| add cross-modal error correction (M3) (ablation) | CMSteg image steganography · JPEG QF 75 | Cap: 1935 -> 3293 | Helped |
| add cross-modal error correction (M3) (ablation) | CMSteg image steganography · G.N. 0.01 | Cap: 1699 -> 2888 | Helped |
| add cross-modal error correction (M3) (ablation) | CMSteg image steganography · Scale 0.5× | Cap: 3725 -> 4174 | Helped |
| add cross-modal error correction (M3) (ablation) | CMSteg image steganography · Scale 2.0× | Cap: 3809 -> 4292 | Helped |
| increase maximum stego-text token length (ablation) | CMSteg · JPEG QF 95 | Cap: 3608 -> 3986 | Helped |
Limitations
- Preceding token errors propagate through the autoregressive model, so recovery rate is not proportional to effective capacity.
- Error-correction capacity is bounded by the carrying capacity of the generated text, and capacity is limited by max text token length.
- Security is argued via a computational reduction and evaluated only against three steganalyzers.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet
Metrics: token recovery rate effective capacity (bits) detection error rate resolution
Baselines: Discop-ImageGPT PARIS StegaStyleGAN
Builds on: Discop PARIS StegaStyleGAN InstructBLIP LlamaGen
Closest papers by meaning
- SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images (ICML 2025 · similarity 0.89)
- StegaStyleGAN: Towards Generic and Practical Generative Image Steganography (AAAI 2024 · similarity 0.87)
- STEAD: Robust Provably Secure Linguistic Steganography with Diffusion Language Model (NeurIPS 2025 · similarity 0.86)
- Robust Message Embedding via Attention Flow-Based Steganography (CVPR 2025 · similarity 0.86)
- StegoGAN: Leveraging Steganography for Non-Bijective Image-to-Image Translation (CVPR 2024 · similarity 0.85)
- On the Robustness of Watermarking for Autoregressive Image Generation (NeurIPS 2026 · 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.