CFR2: LLM-Empowered Device-Free Cross-Subject Feature Recognition Utilizing Channel Frequency Response in Wireless Communication

With the ubiquitous deployment of WiFi infrastructure, device-free human identification leveraging wireless signals has become feasible. However, existing techniques predominantly rely on training data collected from specific users, leading to significant performance degradation when applied to unseen subjects due to inherent biological diversity, environmental noise, and device placement variations. This limitation stems from their inability to generalize physiological or behavioral features across individuals, hindering practical applicability in real-world scenarios. To address this fundamental challenge of Cross-subject Feature Recognition, we propose CFR², a novel framework empowered by Large Language Models (LLMs) utilizing Channel Frequency Response (CFR). CFR² harnesses the powerful generalization and contextual understanding capabilities of LLMs. Specifically, it introduces a Retrieval-Augmented Generation (RAG) pipeline coupled with Reparameterized Parameter-Efficient Fine-Tuning (R-PEFT) tailored for CFR recognition. By leveraging LLMs' ability to learn complex patterns beyond subject-specific training constraints, CFR² significantly improves the generalization capability and practical utility of wireless sensing for feature recognition. Experimental evaluations demonstrate its effectiveness. © 2025 IEEE.