DWI-Alarm: A Device-Free Small Target Indoor Intrusion Detection System Leveraging Diffusion Models

Indoor Intrusion Detection, as a core technology in intelligent security systems, has recently achieved breakthroughs in passive Non-Line-of-Sight (NLoS) sensing by leveraging finegrained Channel State Information (CSI) from commercial WiFi devices. However, when detection targets scale down from humansized objects to small-scale entities, the performance of existing methods degrades significantly due to signal attenuation and feature distortion caused by multimodal obstacles in complex NLoS environments. To address the challenge of small-target intrusion detection in such scenarios, this paper proposes a diffusion model-based CSI feature reconstruction framework. The framework employs a multi-scale denoising network to restore signal feature loss and introduces a Structured State Space Model to dynamically fuse heterogeneous sensitivities across subcarriers. Experimental results demonstrate that the proposed method achieves a $10.2 \%$ reduction in false alarm rates, providing a novel theoretical framework for millimeter-level target perception in highly cluttered NLoS environments.