Seminars
Home
Centers & Programs
AI and Natural Sciences
Seminars
- FIELD
- AI and Natural Sciences
- DATE
-
Nov 19 (Wed), 2025
- TIME
- 15:00 ~ 17:00
- PLACE
- 7323
- SPEAKER
- Yang, Dongyoon
- HOST
- Kim, Dongwon
- INSTITUTE
- Data Intelligence, SK hynix Inc.
- TITLE
- Generalizing Adversarial Robustness: Tight Bound for Multi-Domain Adaptation
- ABSTRACT
- Adversarial robustness under domain adaptation remains a fundamental challenge, particularly when models must generalize across multiple target domains.
In this work, we develop a generalized theoretical framework for multi-target robust domain adaptation, extending the foundation of TAROT to provide a deeper and more unified understanding of robust generalization across domains.
While the original TAROT established an upper bound on the robust risk, we further derive a complementary lower bound with analogous divergence terms, resulting in a tight generalization bound that theoretically characterizes the limits of robust transferability among domains.
This formulation unifies the analysis of robustness and domain invariance under adversarial conditions and provides theoretical insight into their interaction across multiple targets.
Furthermore, we reveal that existing robust unsupervised domain adaptation methods tend to amplify robustness disparities across classes, causing imbalance and degraded transfer robustness.
Our framework offers a principled explanation for this phenomenon and demonstrates, through both theory and experiment, that the extended TAROT approach effectively mitigates these disparities while preserving strong domain-invariant robustness.
Comprehensive evaluations on diverse multi-domain and multi-target benchmarks confirm that our framework delivers improved robustness, stability, and balance, establishing a tighter and more general theoretical foundation for robust domain adaptation.
- FILE
-
