GraphMind: LLM-Driven Botnets Evade Social Bot Detection
A new study from arXiv introduces GraphMind, a framework that enables large language model (LLM)-driven social bots to learn and replicate human-like social network structures. Current LLM-based bots mimic local interactions but fail to coordinate globally, making them detectable by graph neural networks (GNNs). GraphMind addresses this by making bots graph-aware, allowing them to construct realistic network topologies. The researchers then built GraphMind-Botnet to test existing detection algorithms. Experiments showed that both text-based and graph-based detection models performed significantly worse against GraphMind-Botnet, highlighting a new vulnerability in social bot detection. The paper is available on arXiv under identifier 2605.12512.
Key facts
- GraphMind equips LLM-driven social bots to learn human-like social network structures.
- Current LLM-based bots are graph-unaware and vulnerable to GNN-based detection.
- GraphMind-Botnet was built to evaluate detection algorithms.
- Experiments showed degraded performance in both text-based and graph-based detection models.
- The paper is published on arXiv with ID 2605.12512.
- The study focuses on social bot detection evasion.
- LLM-based bots can autonomously engage in local interactions.
- GraphMind enables global coordination over network structures.
Entities
Institutions
- arXiv