Research map / Adversarial robustness and security / Certified and empirical robustness
Multitask Learning Strengthens Adversarial Robustness
Chengzhi Mao, Amogh Gupta, Vikram Nitin, Baishakhi Ray, Shuran Song, Junfeng Yang, Carl Vondrick
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 depth and auto-encoder auxiliary tasks to segmentation training (ablation) | DRN-22 segmentation · Cityscapes PGD100 | mIoU: 10.15 -> 11.52 | Improved |
| add auto-encoder task to segmentation (ablation) | DRN-22 segmentation · Cityscapes Houdini100 | mIoU: 5.04 -> 6.24 | Improved |
| add auxiliary tasks to training (single-task attack) | DRN-22 segmentation · Cityscapes | mIoU / abs error: segmentation mIoU improves by 1.2 points and depth by 11% under attack | Improved |
| add segmentation task to depth training (ablation) | DRN-22 depth · Cityscapes PGD100 | abs error: 10.8 -> 9.69 | Improved |
| add auxiliary tasks to training (ablation) | DRN-22 · Cityscapes clean | mIoU: 48.58 -> 48.61 | No clear effect |
| add auxiliary tasks to training | ResNet-18 on 11 Taskonomy tasks · Taskonomy PGD L_inf=4 50 steps | task performance: 10.23% relative improvement | Improved |
| add auxiliary tasks to training | ResNet-18 on 11 Taskonomy tasks · Taskonomy clean | task performance: average improvement of 4.7% | Improved |
| add multi-task learning to adversarial training | DRN-22 · Cityscapes adversarial training | mIoU / abs error: segmentation mIoU improves by 2.40 points and depth improves by 8.4% | Improved |
| add depth task to adversarially trained segmentation (ablation) | DRN-22 · Cityscapes adversarial training Houdini100 | mIoU: 17.05 -> 19.45 | Improved |
| add auxiliary tasks under multi-task PGD attack (ablation) | DRN-D-105 · Cityscapes PGD | mIoU: 15.75 -> 16.36 | Improved |
| add auxiliary tasks under multi-task PGD attack (ablation) | DRN-D-105 · Cityscapes PGD | depth abs error: 6.81 -> 5.81 | Improved |
| increase number of output dimensions (random pixel subset of segmentation output) (ablation) | DRN-105 · Cityscapes FGSM/PGD | robustness | Improved |
Closest papers by meaning
- Adversarial Robustness in Multi-Task Learning: Promises and Illusions (AAAI 2022 · similarity 0.94)
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.89)
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · similarity 0.89)
- Adversarially Robust Multi-task Representation Learning (NeurIPS 2024 · similarity 0.89)
- Exploring Architectural Ingredients of Adversarially Robust Deep Neural Networks (NeurIPS 2021 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
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.
