From Discrete to Continuous: One-Shot Multi-Position Character-Level Adversarial Attacks

Lingyi Kong, Yutao Yang, Zhanghao Hu, Qilong Qiu, Zheng Wang, Jingjing Xue, Lin Gui, Feiping Nie

Keywords: OSCR-Attack: One-Shot Character Level Attacks through Self-Optimizing Continuous Relaxation

Abstract: 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 beliefs, we introduce two adaptively learnable matrices that transform discrete choices into continuous representations, enabling automatic one-shot multi-position, multi-character insertion. To optimise the two learnable matrices, we propose OSCR-Attack, an end-to-end framework based on gradient-based optimisation, with a conflict resolution strategy mapping the optimised continuous distributions back into discrete insertion operations. Extensive experiments on three benchmarks with three open-source LLMs show that OSCR-Attack improves attack success rate (ASR) by up to 16\% points and accelerates the attack by up to 6 times compared to recent baselines.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 18592