Research map / Label-efficient and robust learning / Domain adaptation and generalization
TAROT: Towards Essentially Domain-Invariant Robustness with Theoretical Justification
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.
Derives a robust margin disparity discrepancy and generalization bound for adversarially robust unsupervised domain adaptation, and proposes TAROT, which combines teacher pseudo-labels, adversarial training, and MDD-style alignment starting from a robustly pretrained model.
Problem
Adversarially robust unsupervised domain adaptation: keeping robust accuracy on an unlabeled target domain under adversarial perturbations, where existing robust UDA methods lack theoretical grounding and generalize poorly to source and unseen domains.
Approach
Extends MDD to a robust margin disparity discrepancy (robust target disparity minus standard source disparity) giving a target robust-risk bound; the resulting TAROT objective combines adversarial cross-entropy on teacher pseudo-labels with robust-margin discrepancy alignment, initialized from a robustly pretrained model.
Key results
- On Office-31 (eps=16/255), TAROT averages 89.31% standard and 83.27% robust accuracy, versus 88.03% / 82.95% for PL.
- On Office-Home (eps=16/255), TAROT averages 68.29% standard and 46.80% robust accuracy, versus 66.00% / 44.38% for PL.
- On DomainNet (eps=4/255), TAROT averages 34.50 / 20.94 (C->I setting averages) and 29.51 / 18.00 across I->* tasks, outperforming PL and other baselines in reported averages.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TAROT instead of PL | Robust UDA · Office-31 Avg. | standard accuracy: 88.03 -> 89.31 | Helped |
| Use TAROT instead of PL | Robust UDA · Office-31 Avg. | robust accuracy (AA): 82.95 -> 83.27 | Helped |
| Use TAROT instead of PL | Robust UDA · Office-Home Avg. | standard accuracy: 66.00 -> 68.29 | Helped |
| Use TAROT instead of PL | Robust UDA · Office-Home Avg. | robust accuracy (AA): 44.38 -> 46.80 | Helped |
| Use TAROT instead of PL | Robust UDA · DomainNet C → I | standard accuracy: 10.59 -> 14.65 | Helped |
| Use TAROT instead of PL | Robust UDA · DomainNet C → I | robust accuracy (AA): 6.69 -> 8.16 | Helped |
| Use TAROT instead of PL | Robust UDA · VisDA2017 Syn. → Real | standard accuracy: 66.71 -> 67.12 | Helped |
| Use TAROT instead of PL | Robust UDA · VisDA2017 Syn. → Real | robust accuracy (AA): 37.83 -> 37.77 | Hurt |
| Add Robust-PT initialization (ablation) | TAROT and PL · Robust UDA tasks across perturbation budgets | accuracy | Helped |
| Vary α in the TAROT objective (ablation) | TAROT · Target and unseen domains across tasks | accuracy | Mixed |
Surprising findings
- Standard MDD training yields exactly 0.0 robust accuracy on Office-31, Office-Home and VisDA-2017 at the larger perturbation budgets used.
- Without Robust-PT initialization, training with larger perturbations (eps=16/255) can lead to non-decreasing training loss on some tasks.
Limitations
- Pseudo-labeling with a fixed teacher can suffer confirmation bias and scalability limits on large or diverse target domains, per the authors.
- Performance depends on the perturbation budget and the alpha trade-off, which varies strongly across tasks (requires per-task tuning).
- Relies on a robustly pretrained initialization; without it, large-epsilon training can fail.
Experimental setup
Datasets: Office-31 Office-Home VisDA2017 DomainNet
Benchmarks: Office-31 Office-Home VisDA2017 DomainNet
Metrics: standard accuracy robust accuracy under AutoAttack
Baselines: ARTUDA RFA SRoUDA PL (pseudo labeling with PGD-AT) MDD
Builds on: MDD (Margin Disparity Discrepancy) DANN PGD adversarial training Robust-PT RFA
Closest papers by meaning
- Bridging Theory and Algorithm for Domain Adaptation (ICML 2019 · similarity 0.87)
- Transferable Adversarial Training: A General Approach to Adapting Deep Classifiers (ICML 2019 · similarity 0.87)
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · similarity 0.86)
- f-Domain Adversarial Learning: Theory and Algorithms (ICML 2021 · similarity 0.86)
- Set Valued Predictions For Robust Domain Generalization (ICML 2025 · similarity 0.86)
- Toward Improving Robustness and Accuracy in Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.85)
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.