Research map / Image generation / Editing and style transfer
StegoGAN: Leveraging Steganography for Non-Bijective Image-to-Image Translation
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 StegoGAN, a CycleGAN extension that makes steganography explicit in feature space via a learned unmatchability mask, so generators avoid hallucinating target-domain classes absent from the source domain.
Problem
Unpaired image-to-image translation between domains with non-bijective class correspondence (unmatchable classes present only in the target), where GANs hallucinate spurious objects and CycleGAN-style models hide information via steganography.
Approach
Split the generator's intermediate features into matchable and unmatchable parts using a learned mask, use the unmatchable part explicitly for reconstruction under noise perturbation, and enforce matchable consistency and sparse/binary mask regularization so that forward translations do not invent unmatchable classes.
Key results
- On PlanIGN, StegoGAN achieves RMSE 22.5 vs 27.0 for CycleGAN and Acc(σ2) 74.8, with FID 58.4.
- On Brats MRI, StegoGAN reduces pFPR to 1.1 vs 21.9 for CycleGAN and 15.1 for SRUNIT.
- On GoogleMaps, StegoGAN keeps a false positive rate of 0 across all highway ratios, while competing methods' rates increase.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use StegoGAN instead of CycleGAN | CycleGAN · PlanIGN | RMSE: 27.0 -> 22.5 | Helped |
| use StegoGAN instead of CycleGAN | CycleGAN · PlanIGN | Acc(σ1): 54.5 -> 66.1 | Helped |
| use StegoGAN instead of CycleGAN | CycleGAN · PlanIGN | Acc(σ2): 57.6 -> 74.8 | Helped |
| use StegoGAN instead of the strongest baseline | PlanIGN translation · PlanIGN | FID: 60.2 -> 58.4 | Helped |
| use StegoGAN instead of the strongest baseline | PlanIGN translation · PlanIGN | KID: 2.2 -> 2.4 | Hurt |
| use StegoGAN instead of the strongest baseline | Brats MRI Flair → T1 translation · Brats MRI Flair → T1 | RMSE: 39.8 -> 36.3 | Helped |
| use StegoGAN instead of the strongest baseline | Brats MRI Flair → T1 translation · Brats MRI Flair → T1 | pFPR(‱): 15.1 -> 1.1 | Helped |
| use StegoGAN instead of the strongest baseline | Brats MRI Flair → T1 translation · Brats MRI Flair → T1 | iFPR: 21.8 -> 4.2 | Helped |
| use StegoGAN instead of the strongest baseline | Brats MRI Flair → T1 translation · Brats MRI Flair → T1 | FID: 58.9 -> 58.5 | Helped |
| use StegoGAN instead of the strongest baseline | Brats MRI Flair → T1 translation · Brats MRI Flair → T1 | KID: 2.9 -> 2.4 | Helped |
| vary encoder depth from the full method's depth 1 (ablation) | StegoGAN · PlanIGN | RMSE: 22.5 -> 22.4 (depth -1); 22.5 -> 22.8 (depth 3); 22.5 -> 24.2 (depth 5); 22.5 -> 24.5 (depth 8) | Mixed |
| add Lmatch (ablation) | StegoGAN · GoogleMaps | RMSE: 24.1 -> 22.7 | Helped |
Surprising findings
- Steganography in CycleGAN is shown to be exploitable: the paper turns it into a mechanism to detect and prevent spurious generations rather than suppressing it.
Limitations
- Unmatchability is subjective and requires tuning the lambda_reg hyperparameter to balance hallucination removal against loss of necessary detail.
- Adds complexity to CycleGAN training with an extra mask module and two hyperparameters.
- Only GAN-based; not extended to diffusion models.
Experimental setup
Datasets: PlanIGN GoogleMaps Brats2018
Benchmarks: PlanIGN GoogleMaps Brats MRI
Metrics: RMSE Acc(σ1) Acc(σ2) FID KID pFPR iFPR
Baselines: CycleGAN DRIT GcGAN CUT SRUNIT
Builds on: CycleGAN CUT GcGAN DRIT SRUNIT
Closest papers by meaning
- StegaStyleGAN: Towards Generic and Practical Generative Image Steganography (AAAI 2024 · similarity 0.87)
- TraVeLGAN: Image-To-Image Translation by Transformation Vector Learning (CVPR 2019 · similarity 0.87)
- Generating steganographic images via adversarial training (NeurIPS 2017 · similarity 0.86)
- Provably Secure Image Robust Steganography via Cross-modal Error Correction (AAAI 2025 · similarity 0.85)
- Rethinking the Paradigm of Content Constraints in Unpaired Image-to-Image Translation (AAAI 2024 · similarity 0.85)
- MDDM: Practical Message-Driven Generative Image Steganography Based on Diffusion Models (ICML 2025 · 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.