堅牢性は転移可能なスキル:VLAポリシーのタスク横断的敵対的防御
Robustness as a Transferable Skill: Cross-Task Adversarial Defense for Vision-Language-Action Policies
VLAポリシーに対する敵対的防御を、ソースタスクで学習した堅牢性としてターゲットタスクへ転移させる手法SN-TRVを提案し、実機PiPERで敵対的パッチ下の成功率を20.3%から45.8%に改善した。
著者: Jinquan Zhang, Run Yang, Dongfu Yin, Qiang Lai, Yufeng Yan, Pengshuai Yin, F. Richard Yu
分類: cs.RO, cs.CV
原文アブストラクト
Vision-Language-Action (VLA) policies are vulnerable to adversarial visual perturbations, while defenses often require task-specific adversarial fine-tuning. We investigate whether robustness learned on source manipulation tasks can protect independently specialized target policies without target-side adversarial training. We first show that even a single source defense update improves robustness on held-out tasks, establishing cross-task defense reuse independently of multi-source aggregation. To combine source updates, we introduce Source-Norm Transferable Robustness Vector (SN-TRV), which retains the direction of their mean while calibrating its magnitude using source-update norms. The resulting update is applied to the target visual encoder without target adversarial data, target-side gradient optimization, or additional inference modules. Evaluations across three VLA backbones, fixed and adaptive attacks, and different source defense recipes show improved robustness over naive averaging and identify limits in attack coverage. Physical PiPER trials increase average success under printed adversarial patches from 20.3% to 45.8%.
