一作ACL 2026Findings· Findings of ACL
OSCR -Attack: One-Shot Character Level Attacks through Self-Optimizing Continuous Relaxation
Lingyi Kong · Zhuo Liu · Zhanghao Hu · Qilong Qiu · Yutao Yang · Jingjing Xue · Zheng Wang · Lin Gui · Feiping Nie
摘要
Adversarial attacks have attracted growing attention across domains, including natural language processing (NLP). Character-level adversarial attacks preserve semantics, but they have received less attention because the discrete operations they use are costly and inefficient. Challenging these be...