Research map / Label-efficient and robust learning / Domain adaptation and generalization

Toward Improving Robustness and Accuracy in Unsupervised Domain Adaptation

AAAI 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Feature alignment for unsupervised adaptation

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.

Proposes CAM+SPLR, a two-step SGD self-training method for adversarially robust unsupervised domain adaptation that aligns attention maps of clean and adversarial target examples and refines pseudo labels.

Problem

Adversarial robustness and clean accuracy in unsupervised domain adaptation are hard to achieve because target labels are unavailable and pseudo labels are noisy, which harms adversarial self-training.

Approach

Initialize a frozen Anchor model and a trainable TargetNet from a DANN pre-trained UDA model; TargetNet is trained on PGD adversarial versions of pseudo-labeled target data with an attention-consistency loss against the Anchor's clean-image attention maps, followed by a second SGD step that refines pseudo labels using the change in source cross-entropy as feedback.

Key results

• On PACS, average accuracy improves about 6% and average robustness about 10.2% over state-of-the-art methods at epsilon 2/255 (as reported in the abstract).
• On VisDA, the method improves average accuracy by 8.1% and robustness by 4.9% relative to state-of-the-art baselines.
• On VisDA Syn->Real under PGD-20 at epsilon 8/255, the method reaches 69.5 clean and 57.8 robust accuracy versus DART's 65.6 and 51.2.

Measured effects reported in the paper

Limitations

• Performance depends on the choice of UDA baseline; a weaker baseline lowers robustness and accuracy.
• Robustness drops to near zero at large perturbation sizes for all methods, including the proposed one.

Experimental setup

Datasets: OfficeHome PACS VisDA

Benchmarks: OfficeHome PACS VisDA

Metrics: clean accuracy robust accuracy under PGD-20

Baselines: DANN UDA+AT UDA+Trades UDA+Mart ARTUDA SRoUDA DART

Builds on: DANN TRADES MART ARTUDA DART

Closest papers by meaning

• SRoUDA: Meta Self-Training for Robust Unsupervised Domain Adaptation (AAAI 2023 · similarity 0.92)
• Can We Evaluate Domain Adaptation Models Without Target-Domain Labels? (ICLR 2024 · similarity 0.90)
• Guiding Pseudo-Labels With Uncertainty Estimation for Source-Free Unsupervised Domain Adaptation (CVPR 2023 · similarity 0.90)
• Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.88)
• CASUAL: Conditional Support Alignment for Domain Adaptation with Label Shift (AAAI 2025 · similarity 0.88)
• On Unsupervised Domain Adaptation: Pseudo Label Guided Mixup for Adversarial Prompt Tuning (AAAI 2024 · similarity 0.88)