Research map / Adversarial robustness and security / Certified and empirical robustness
GAT: Guided Adversarial Training with Pareto-optimal Auxiliary Tasks
Salah GHAMIZI, Jingfeng Zhang, Maxime Cordy, Mike Papadakis, Masashi Sugiyama, YVES LE TRAON
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 Macro domain-knowledge auxiliary task with MGDA weighting | adversarial training on CIFAR-10 (ResNet-50) · CIFAR-10 full data | robust accuracy: 39.09 -> 48.38 | Improved |
| add Jigsaw or Rotation auxiliary task with MGDA weighting | adversarial training on CIFAR-10 · CIFAR-10 full data | robust accuracy: Jigsaw 32.95, Rotation 36.13 vs 39.09 baseline | Hurt |
| remove curvature regularization from GAT (ablation) | GAT Macro on CIFAR-10 · CIFAR-10 full data | robust accuracy: 48.38 -> 44.34 robust; clean 73.70 -> 83.47 | Hurt |
| add Macro auxiliary task in limited-data training | adversarial fine-tuning on 10% CIFAR-10 · CIFAR-10 10% data | robust accuracy: 8.37 -> 22.42 | Improved |
| add Jigsaw auxiliary task on 10% data | adversarial fine-tuning on CIFAR-10 10% · CIFAR-10 10% data | robust accuracy: 8.37 -> 32.41 (PGD-10) | Improved |
| add Jigsaw auxiliary task on CIFAR-10 at 50% data (ablation) | adversarial training on CIFAR-10 50% · CIFAR-10 50% data | robust accuracy: 33.69 -> 23.79 (CIFAR-10 table, Jigsaw) | Hurt |
| add Rotation auxiliary task on CIFAR-10 at 50% data (ablation) | adversarial training on CIFAR-10 50% · CIFAR-10 50% data | robust accuracy: 33.69 -> 16.63 | Hurt |
| add Gender auxiliary task (CheXpert) | single-task AT for Atelectasis · CheXpert Atelectasis | robust AUC: 50.00 -> 83.00 | Improved |
| add Age auxiliary task (CheXpert) (ablation) | single-task AT for Atelectasis · CheXpert Atelectasis | robust AUC: 50.00 -> 46.26 | Hurt |
| add Jigsaw auxiliary task (CheXpert) | single-task AT for Edema · CheXpert Edema | robust AUC: 55.69 -> 70.47 | Improved |
| add Gender auxiliary task (CheXpert Edema) (ablation) | single-task AT for Edema · CheXpert Edema | robust AUC: 55.69 -> 31.46 | Hurt |
| add Macro auxiliary task (CIFAR-10 at 10% under PGD-4 evaluation) (ablation) | limited-data adversarial fine-tuning · CIFAR-10 10% data | robust accuracy: 11.81 -> 42.68 (PGD-4) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MaxUp — MaxUp: Lightweight Adversarial Training With Data Augmentation Improves Neural Network Training (CVPR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Guided Adversarial Attack for Evaluating and Enhancing Adversarial Defenses (NeurIPS 2020 · similarity 0.87)
- Efficient and Effective Augmentation Strategy for Adversarial Training (NeurIPS 2022 · similarity 0.86)
- Mixture of Adversarial LoRAs: Boosting Robust Generalization in Meta-Tuning (NeurIPS 2024 · similarity 0.84)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.84)
- Boosting Adversarial Robustness with CLAT: Criticality Leveraged Adversarial Training (ICML 2025 · similarity 0.84)
- Taxonomy Driven Fast Adversarial Training (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.
