"Wireless Backdoor Attack against Large Language Model Empowered Channel Prediction"

"The integration of Large Language Models (LLMs) into wireless communications for channel state information (CSI) prediction introduces transformative capabilities but also exposes critical security vulnerabilities, particularly backdoor attacks. This paper investigates how adversaries exploit the openness of wireless propagation where signals are inherently susceptible to eavesdropping and adversarial interference, and the black-box nature of neural networks to inject stealthy triggers (e.g., Gaussian white, narrowband, or impulse interference) into online training samples. These triggers corrupt CSI estimates during the model dynamic adaptation to wireless environments, causing LLMs to produce maliciously altered predictions while maintaining normal behavior without trigger corruption. We demonstrate that narrowband interference, due to its frequency-specific and covert characteristics, serves as the most effective trigger. To counter these threats, we propose two defense strategies: physical-layer authentication using VT-CNN2-based channel fingerprinting achieves high accuracy in distinguishing adversarial signals.And adversarial training that enhances model robustness by exposing LLMs to adversarial samples during online updates. Simulations using 3GPP-compliant datasets validate the attack efficacy, showing significant increases in normalized mean square error (NMSE) degradation under narrowband triggers, while defense methods reduce NMSE to near-clean conditions. "