Research map / Label-efficient and robust learning / Domain adaptation and generalization
Understanding the Limits of Unsupervised Domain Adaptation via Data Poisoning
Akshay Mehra, Bhavya Kailkhura, Pin-Yu Chen, Jihun Hamm
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 mislabeled target-domain data to the source training data | DANN · SVHN → MNIST | target domain accuracy: 78.05$\pm$1.15 -> {\bf 1.46$\pm$1.12} | Hurt |
| Add mislabeled target-domain data to the source training data | CDAN · MNIST → MNIST_M | target domain accuracy: 73.88$\pm$1.10 -> {\bf0.59$\pm$0.12} | Hurt |
| Add mislabeled target-domain data to the source training data | MCD · MNIST → USPS | target domain accuracy: 89.96$\pm$2.04 -> {\bf0.66$\pm$0.16} | Hurt |
| Add mislabeled target-domain data to the source training data | SSL · USPS → MNIST | target domain accuracy: 82.23$\pm$1.59 -> {\bf9.88$\pm$1.07} | Hurt |
| Add mislabeled target-domain data to the source training data | DANN · A → D | target domain accuracy: 84.06 -> {\bf59.83$\pm$0.20} | Hurt |
| Add mislabeled target-domain data to the source training data | CDAN · D → A | target domain accuracy: 71.25 -> {\bf11.19$\pm$1.47} | Hurt |
| Use mislabeled source-domain data rather than target-domain data as poisons (ablation) | DANN · SVHN → MNIST | target domain accuracy: 78.05$\pm$1.15 -> 70.26$\pm$2.84 | Hurt |
| Use source-looking watermarked poison data and increase the watermarking percentage (ablation) | DANN · MNIST → USPS | target domain accuracy: 92.17$\pm$0.73 -> 39.29$\pm$1.34$_{0.30}$ | Hurt |
| Use source-domain rather than target-domain base data to initialize clean-label poisons (ablation) | UDA methods · MNIST → MNIST_M, two-class (3 vs 8) | attack success rate | Hurt |
Closest papers by meaning
- Divergence-Agnostic Unsupervised Domain Adaptation by Adversarial Attacks (TPAMI 2021 · similarity 0.93)
- Unsupervised Robust Domain Adaptation: Paradigm, Theory and Algorithm (IJCV 2025 · similarity 0.91)
- Can We Evaluate Domain Adaptation Models Without Target-Domain Labels? (ICLR 2024 · similarity 0.91)
- Cross-Domain Gradient Discrepancy Minimization for Unsupervised Domain Adaptation (CVPR 2021 · similarity 0.90)
- Unsupervised Multi-Source Domain Adaptation Without Access to Source Data (CVPR 2021 · similarity 0.90)
- Distributionally Robust Classification for Multi-source Unsupervised Domain Adaptation (ICLR 2026 · similarity 0.89)
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.
