S-GAP: Stealthy Attacks on GraphLLMs via Unified Primitives
In brief
- S-GAP merges textual and structural perturbations into a single semantic primitive for GraphLLMs.
- Iterative method with surrogate evaluation refines attacks while maintaining stealth.
- Boosts attack success rate by 7% on average across four datasets.
Researchers introduce S-GAP, a unified adversarial attack method for Graph Large Language Models (GraphLLMs). It merges textual and structural perturbations into a single coherent semantic representation, boosting attack success while remaining undetectable. Tests across four datasets show a 7% average increase in success rate over existing methods.
What it changes for you
- CybersecurityGraphLLMs face more effective and stealthier attacks, requiring stronger countermeasures.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error