Research map / Adversarial robustness and security / Attacks on text and graphs
A²RNet: Adversarial Attack Resilient Network for Robust Infrared and Visible Image Fusion
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 A²RNet, a U-Net with a transformer-style defensive refinement module trained adversarially (PGD attacks plus an anti-attack loss using pseudo-labels) to make infrared-visible image fusion robust to adversarial perturbations.
Problem
Infrared and visible image fusion (IVIF) methods are fragile under adversarial perturbations, producing artifacts in fused images and degrading downstream detection and segmentation, and there is no fusion-oriented adversarial training paradigm (no ground truth for fusion).
Approach
An adversarial training paradigm for IVIF that generates PGD attacks on both input modalities guided by pseudo-labels from a common CNN, combined with a U-Net pipeline whose middle includes a transformer-based defensive refinement module with a Mercer-kernel linear-complexity self-attention.
Key results
- Under adversarial attack, A²RNet reaches mAP@.5 of 0.781 (drop of 0.024) and mIoU of 0.677 on the downstream tasks, the best among compared methods.
- Ablation on MFNet: the full model (U-Net + DRM) achieves mIoU 0.745 versus 0.594 without U-Net and 0.669 without DRM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use A²RNet instead of SeAFusion | M³FD detection · M³FD adversarial | mAP@.5: 0.732 -> 0.781 | Helped |
| Use A²RNet instead of EMMA | MFNet segmentation · MFNet adversarial | mIoU: 0.649 -> 0.677 | Helped |
| Add U-Net to the architecture (ablation) | A²RNet · MFNet | mIoU: 0.594 -> 0.745 | Helped |
| Add U-Net to the architecture (ablation) | A²RNet · MFNet | EN: 5.231 -> 6.543 | Helped |
| Add DRM to the architecture (ablation) | A²RNet · MFNet | mIoU: 0.669 -> 0.745 | Helped |
| Add DRM to the architecture (ablation) | A²RNet · MFNet | CC: 0.587 -> 0.605 | Helped |
| Train with adversarial training rather than non-AT (ablation) | A²RNet · Adversarial fusion and downstream tasks | performance | Helped |
| Use the anti-attack loss rather than the ablation loss (ablation) | A²RNet · Adversarial fusion | fusion robustness | Helped |
| Apply the anti-attack loss and pseudo-labels to PAIF (ablation) | PAIF · Adversarial fusion | texture features and target details | No clear effect |
Surprising findings
- The anti-attack loss is reported not to be plug-and-play: embedding it into PAIF did not noticeably improve its robustness.
Limitations
- Pseudo-labels come from a self-trained common CNN rather than SOTA fusion outputs, which may limit the fusion quality ceiling.
- Attacks are only PGD-based with a small set of perturbation settings (epsilon 4/255), so generality to stronger or other attacks is not shown.
- Ablation of the anti-attack loss is limited to one transplant experiment on PAIF.
Experimental setup
Datasets: M3FD MFNet
Benchmarks: M3FD MFNet
Metrics: EN SD PSNR CC SCD mAP@.5 mIoU
Baselines: TarDAL SeAFusion IGNet PAIF CoCoNet LRRNet EMMA
Builds on: PGD U-Net PAIF FGSM
Closest papers by meaning
- Fusion Is Not Enough: Single Modal Attacks on Fusion Models for 3D Object Detection (ICLR 2024 · similarity 0.86)
- Advancing Adversarial Robustness in GNeRFs: The IL2-NeRF Attack (CVPR 2025 · similarity 0.85)
- Dispel Darkness for Better Fusion: A Controllable Visual Enhancer based on Cross-modal Conditional Adversarial Learning (CVPR 2024 · similarity 0.84)
- Random Entangled Tokens for Adversarially Robust Vision Transformer (CVPR 2024 · similarity 0.84)
- Harnessing the Computation Redundancy in ViTs to Boost Adversarial Transferability (NeurIPS 2025 · similarity 0.84)
- NeRFail: Neural Radiance Fields-Based Multiview Adversarial Attack (AAAI 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.