Research map / Adversarial robustness and security / Attacks on text and graphs
Diffusion-Based Native Adversarial Synthesis for Enhanced Medical Segmentation Generalization
Hongyu Zhang, Haipeng Chen, Zhimin Xu, Chengxin Yang, Yingda Lyu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adversariality miner (noise reselection) | SiameseDiff synthetic augmentation · ACDC (nnU-Net) | DSC gain: 3.06 -> 8.14 | Improved |
| add adversariality miner | DiffBoost · Synapse (nnU-Net) | DSC gain: 4.09 -> 9.12 | Improved |
| add adversariality miner | SiameseDiff · Polyps device shift (ETIS) | DSC gain: 2.5 -> 7.8 | Improved |
| add adversariality miner | DiffBoost · MMWHS CT->MRI | DSC gain: 15.8 -> 25.6 | Improved |
| keep only high-adversariality synthetic samples (loss > 0.4) (ablation) | Utrain + synthetic · Polyps (nnU-Net) | per-sample gain: 4.48e-3 -> 39.19e-3 per sample | Improved |
| add attack-based adversarial guidance (artificial adversariality) (ablation) | SiameseDiff synthesis · Polyps | generalization gain | Hurt |
| add attack-based guidance (ablation) | SiameseDiff synthesis · Polyps | FID: 65.21 -> 102.32 (AdvDiffuser) | Hurt |
| add DiffAug perturbation augmentation | DiffBoost/SiameseDiff · Polyps (nnU-Net) | DSC gain: 5.05 -> -3.22 | Hurt |
| increase synthetic budget beyond 3x N (ablation) | base sampling · Polyps/Synapse | DSC gain | Hurt |
| increase synthetic budget with adversariality miner (ablation) | Ours · Polyps/Synapse | DSC gain: 9.09 at 1xN to 14.30 at 6xN (Synapse) | Improved |
| use too small or too large KL weight beta (ablation) | Adversariality miner · Polyps | DSC gain | Hurt |
| optimize through more than 25 DDIM steps (ablation) | Adversariality miner · Polyps | DSC gain | Hurt |
Closest papers by meaning
- Structure-Guided Adversarial Training of Diffusion Models (CVPR 2024 · similarity 0.88)
- MimicDiffusion: Purifying Adversarial Perturbation via Mimicking Clean Diffusion Model (CVPR 2024 · similarity 0.88)
- Noise-Consistent Siamese-Diffusion for Medical Image Synthesis and Segmentation (CVPR 2025 · similarity 0.88)
- CAT: Contrastive Adversarial Training for Evaluating the Robustness of Protective Perturbations in Latent Diffusion Models (ICML 2025 · similarity 0.88)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.88)
- AdvAD: Exploring Non-Parametric Diffusion for Imperceptible Adversarial Attacks (NeurIPS 2024 · similarity 0.87)
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.
