Research map / Adversarial robustness and security / Black-box and transfer attacks
Stealthy Multi-task Adversarial Attacks
Jiacheng Guo, Tianyun Zhang, Lei Li, Haochen Yang, Hongkai Yu, Minghai Qin
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 negative weights on non-targeted task losses (stealthy formulation) (ablation) | PGD attack · NYUv2 (PGD Linf, eps=4/255, target depth) | segment mIoU on non-targeted task: 24.15 -> 47.67 (clean 46.56) | Improved |
| add negative weights on non-targeted task losses (stealthy formulation) (ablation) | PGD attack · Cityscapes (PGD Linf, eps=4/255, target part seg) | segment mIoU on non-targeted task: 33.98 -> 54.36 (clean 54.20) | Improved |
| use stealthy weighting for the targeted task on NYUv2 (ablation) | PGD attack · NYUv2 (PGD Linf, eps=4/255, target segmentation) | segment mIoU on targeted task: 17.04 -> 18.33 | No clear effect |
| automated weight search instead of manual search (ablation) | SMTA2 · NYUv2 (PGD Linf, eps=4/255) | targeted-task metric: segment 18.30 vs 18.33; depth 165.27 vs 168.69 | No clear effect |
| use stealthy weighting on Cityscapes disparity target | PGD attack · Cityscapes (PGD L2, eps=5, target disparity) | disparity aErr on targeted task: 921.67 -> 607.56 | Hurt |
| use stealthy weighting on NYUv2 normal target | PGD attack · NYUv2 (PGD L2, eps=10, target normal) | normal mDist on targeted task: 81.43 -> 68.21 | Hurt |
| use stealthy weighting on Cityscapes disparity target (Linf) | PGD attack · Cityscapes (PGD Linf, eps=4/255, target disparity) | disparity aErr on targeted task: 490.15 -> 644.12 | Improved |
| attack an adversarially trained model | non-stealthy PGD · NYUv2 (PGD Linf, eps=4/255, target segmentation) | segment mIoU on targeted task: 17.04 -> 27.37 (undefended vs AT) | Hurt |
| attack an adversarially trained model | SMTA2 Auto · Cityscapes (PGD Linf, eps=4/255, target segmentation) | segment mIoU on targeted task: 17.08 -> 31.22 (undefended vs AT) | Hurt |
Closest papers by meaning
- Adversarially Robust Multi-task Representation Learning (NeurIPS 2024 · similarity 0.84)
- Focus on Hiders: Exploring Hidden Threats for Enhancing Adversarial Training (CVPR 2024 · similarity 0.84)
- Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training (CVPR 2025 · similarity 0.84)
- Adversarial Robustness in Two-Stage Learning-to-Defer: Algorithms and Guarantees (ICML 2025 · similarity 0.83)
- Strong Transferable Adversarial Attacks via Ensembled Asymptotically Normal Distribution Learning (CVPR 2024 · similarity 0.83)
- SuperDeepFool: a new fast and accurate minimal adversarial attack (NeurIPS 2024 · similarity 0.83)
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.
