LLM-Powered Multi-Agent Attacks on Cooperative Spectrum Sensing
Cooperative Spectrum Sensing (CSS) is a promising approach allowing secondary users to utilize unused spectrum without interfering with primary users. However, the collaborative nature of CSS makes it vulnerable to malicious nodes that inject falsified sensing data. Existing attacks have incorporated AI to enhance the effectiveness and overcome challenges in opaque-box settings. However, conventional machine learning models often fail to adapt to rapidly changing wireless environments. In this paper, we propose a large language model (LLM)-powered multi-agent attack framework that leverages the reasoning and adaptability of LLMs to coordinate multiple agents for generating adaptive and context-aware fake sensing reports. The proposed architecture consists of a generator and a discriminator, which work collaboratively to progressively refine falsified sensing data. In addition, to improve prompting efficiency, we incorporate wireless-domain knowledge into the generator to enable advanced prompting, thereby ensuring more effective and efficient malicious report generation. We implement the proposed attack framework in a TV white space system, and experimental results show that our method achieves up to 98.35% system disruption against existing defense mechanisms while requiring significantly fewer feature modifications.