A Novel Two-Stage Attacks on Korean Language Models: Single-Token Triggers Search and Morphology-Preserving Minimal Edits

Key Points

• The aim is to propose a two-stage attack framework targeting Korean language models, leveraging their unique properties.
• Initial stage involves universal adversarial trigger search using gradient information for single-token attacks.
• Second stage targets samples failed in the first stage with morphology-preserving minimal edits.
• Evaluated using NSMC dataset employing KoBERT and KoELECTRA models.
• High attack success rate achieved, with KoBERT at 0.963 and KoELECTRA at 0.940.
• Effective use of triggers placed at the end of sentences due to Korean sentence structure.
• Words that express emotions indirectly functioned as strong triggers.