Abstract
Device-free passive (DfP) intrusion detection systems enable the monitoring of moving entities without requiring them to carry any electronic devices. However, traditional DfP systems often require specific Access Point (AP) deployments to maintain performance, limiting their practical application. This paper introduces IDSDL, an intrusion detection system that leverages finer-grained Channel State Information (CSI) and deep learning to remove these placement constraints. By decomposing CSI phase data into individual propagation path components and utilizing a Convolutional Neural Network (CNN), IDSDL achieves high sensitivity and reliability in complex environments.
The Challenge of Device-Free Detection
Traditional DfP systems primarily rely on Received Signal Strength (RSS) or CSI amplitude. While effective in Line-of-Sight (LOS) scenarios—where an intruder directly crosses the path between a transmitter (TX) and receiver (RX)—these metrics struggle in Non-Line-of-Sight (NLOS) conditions. In such cases, the signal changes caused by an intruder in a corner or behind an obstacle are often too subtle to be detected against the background noise of the dominant direct signal.
To overcome this, researchers have previously resorted to meticulous AP deployment, which is often impractical for home or office security. IDSDL addresses this by focusing on the CSI phase, which is significantly more sensitive to environmental changes than amplitude.
Methodology: The IDSDL Pipeline
The IDSDL system processes WiFi signals through a multi-stage pipeline designed to extract and classify motion-sensitive features automatically.
1. Data Preprocessing
Raw CSI data is inherently noisy due to hardware imperfections and environmental interference. IDSDL employs "phase sanitization" via linear transformation to remove random phase offsets. Additionally, Hampel or Kalman filters are applied to eliminate outliers, ensuring a clean signal for further analysis.
2. Path Decomposition
The core innovation of IDSDL lies in its path decomposition algorithm. By treating the CSI signal as a Vandermonde matrix, the system decomposes the total phase into components representing distinct propagation paths. This allows the system to isolate changes occurring on indirect, reflected paths that would otherwise be masked by the stronger direct path signal.
3. CNN-Based Classification
Instead of manually engineering features, IDSDL converts the decomposed signal into a 2D "feature image." This image is fed into a CNN consisting of three convolutional layers and three dense layers. The CNN learns to distinguish between "static" (no intrusion) and "dynamic" (intrusion) states by recognizing patterns in the signal's waveform and timing.
Experimental Evaluation
The system was prototyped using a TP-LINK router as the transmitter and a computer equipped with an Intel WiFi Link 5300 NIC as the receiver. Testing was conducted in a 56-square-meter laboratory filled with furniture and electronics to simulate a challenging multipath environment.
Key Results
- High Accuracy: IDSDL achieved an average detection accuracy of 98.69% for single participants and 98.91% for multiple participants.
- NLOS Robustness: The system maintained high sensitivity even when intruders were not in the direct line-of-sight, outperforming traditional amplitude-based methods like PADS.
- Real-Time Capability: With a processing latency of approximately 10 ms per sample, IDSDL is well-suited for real-time security monitoring.
Conclusion
IDSDL represents a significant step forward for device-free intrusion detection. By combining sensitive CSI phase decomposition with the automated feature extraction capabilities of deep learning, the system provides a robust, flexible solution that works effectively without the need for specialized hardware placement. Future work will focus on expanding the system's capabilities to classify specific types of activities and optimizing the path decomposition model for even more diverse environments.
Summary updated