[ECCV 2026] Learning from Adversity: Semantic-Aware Mask Refinement through Adversarial Perturbation