1. Introduction
Device-free human detection has attracted a lot of interest in recent years. It can detect human presence in the monitoring area without any sensing-related devices attached to the people [
1]. It can be used well in intrusion detection systems, which is a vital security component in a smart home. Aiming at handling the security issues in a smart home, many techniques have been utilized to implement device-free human detection, such as video-based, infrared-based, Radio Frequency Identification (RFID)-based and Ultra-Wide Bandwidth (UWB)-based approaches. Although they have a good detection accuracy, these approaches have limited using conditions and need dedicated devices that hinder their adoption. WiFi-enabled devices become the catalyst of device-free sensing as they have been widely deployed in both public and private buildings. Besides being used for communication, WiFi networks can also be used as sensor networks [
2,
3,
4]. Many applications have emerged based on WiFi infrastructures, human detection [
5], indoor localization [
6], and even human identification [
7] are some representative applications.
A typical WiFi-based device-free human detection system usually contains several pairs of transmitters and receivers. A wireless router can act as a transmitter, while a WiFi-enabled device can act as a receiver. As a result, it doesn’t have the problem of key management [
8,
9] compared with sensor-based approaches. The rational of WiFi-based device-free human detection is that human presence has an impact on signal propagation, which will cause the signal strength fluctuation at the receiver [
10]. Previous WiFi-based human detection systems utilize Received Signal Strength Indicator (RSSI) from Media Access Control (MAC) layer for it is easy to obtain. However, RSSI is a coarse-grained measurement. In the typical indoor scenario, RSSI becomes unreliable due to multipath fading. It may increase, decrease, or even remain the same when a person moves in the monitoring area. Recently, many studies explore CSI from physical layer of wireless networks to detect human motion [
11,
12,
13]. As indicated in [
14], CSI is a subcarrier-level measurement that is more fine-grained compared with RSSI. It is more sensitive to environmental changes while keeps quite stable in static scenarios. As a result, CSI succeeds in improving the performance of human detection.
However, state-of-the-art human detection techniques still have limitations for intrusion detection systems. Common human detection techniques can only detect a human who is walking with a regular pattern. Nevertheless, an intruder in the building is likely to keep away from the security devices or move very slowly to hide himself from being monitored. Furthermore, most human detection techniques require on-site calibration of both static and dynamic environments. On-site calibration is labor intensive and it needs professional deployment and maintenance that makes a human detection system more complex in practical use. Consequently, human detection techniques will fail in detecting intruders in security systems, and we need to explore effective features to model human motion.
To deal with the limitations, in this work, we propose a Robust Device-Free Intrusion Detection (RDFID) system leveraging fine-grained CSI. We investigate the impact of human motion on WiFi signals and demonstrate that different patterns of human motion in different scenarios can be modeled by a unified framework. First, we extract the wavelet variance of CSIs from frequency domain as the feature. It is more sensitive to human motion, and more robust under different moving patterns. In addition, the feature values of static and intrusion can be seen to be generated by different Gaussian Models. As a result, intrusion can be detected using a Gaussian Mixture Model (GMM). As shown in
Figure 1, RDFID can detect human motion of different moving patterns. In addition, it can be easily deployed that it can achieve a satisfying performance even using a single pair of transceivers, and needs no re-calibration in different scenarios.
We prototyped RDFID in three typical home and office scenarios with commodity WiFi devices composing only one wireless link. We evaluate the system and compare the performance with Fine-grained Real-time passive human motion Detection (FRID), device-free Passive Detection of moving humans with dynamic Speed (PADS) and Fine-grained Indoor Motion Detection (FIMD). The results show that the detection precision of RDFID can achieve over 97% under different moving patterns. Consequently, it makes intrusion detection systems a step closer to practical use.
In summary, the contributions of our work are as follows:
We propose RDFID, a novel device-free WiFi-based intrusion detection approach, which can detect intruders with different moving patterns at a high accuracy, and needs no re-calibration in different scenarios. It can be deployed in smart home scenarios to ensure security.
We extract real-time features from CSIs in frequency domain, which is more sensitive to human motion of various moving patterns.
We use the Gaussian Mixture Model (GMM) as the classifier based on the observation that the feature values under different moving patterns and different environments can be seen to be generated by different Gaussian Models.
In the rest of this paper, the related works about WiFi-based human detection are reviewed in
Section 2. Some preliminaries are introduced in
Section 3.
Section 4 presents the design details of our proposed intrusion detection system, while the performance evaluation is provided in
Section 5. In
Section 6, the potentials and limitations are discussed and we conclude this work in
Section 7.
2. Related Work
WiFi-based passive human detection is the fundamental technique of various ubiquitous wireless sensing applications, such as indoor localization, human identification and activity recognition. It can be widely deployed in smart home scenarios to ensure the security. A large quantity of studies about wireless sensing promote the development of wireless sensing.
Earlier passive human detection systems usually utilize RSSI from the MAC layer of the wireless network. After Youssef et al. proposed the concept of device-free passive human motion detection, they optimized their approach and made the system work in real environments [
10]. Nuzzer leveraged probabilistic techniques, and had the capability to both localize a single entity and estimate the number of people in the area of interest [
15]. Since RSSI is a coarse-grained measurement of wireless networks, many RSSI-based human detection systems deployed multiple pairs of transceivers to achieve a higher accuracy [
16]. Another technique of human detection using multiple pairs of transceivers is Radio Tomographic Imaging (RTI) [
17]. Researchers also developed various approaches based on RTI, such as the kRTI [
18] and dRTI [
19]. However, RSSI-based human detection systems suffer from severe multi-path efficiency [
20]. As a result, more and more researchers move their attention to the more fine-grained measurement, CSI.
To overcome the shortcomings of RSSI-based human detection systems, Fine-grained device-free Motion Detection (FIMD) utilized the burst pattern of CSIs during human motion to detection human presence [
21]. Fine-grained Indoor Localization (FILA) explored the frequency diversity of the subcarriers in Orthogonal Frequency Division Multiplexing (OFDM) systems, and constructed a signal propagation model [
22,
23]. As human motion can cause the fluctuation of the signal, Bfp harnessed the variance of the amplitude of the CSIs to improve the performance of human detection [
11]. PADS took advantages of the whole information of CSI including both amplitude and phase feature to detect human motion with various speeds [
24]. It calculates the maximum eigenvalue of covariance matrix of normalized amplitude and phase information, respectively, as the feature. Support Vector Machine (SVM) is used as the classifier. FRID explored the phase feature of CSIs and achieved calibration-free human detection without the need of a normal profile [
25,
26]. Short-term averaged variance ratio (SVR) and long-term averaged variance ratio which are two schemes based on the coefficient of variance of phase are introduced to eliminate the re-calibration cost. Conventional human detection systems demonstrated directional monitoring coverage, and Zimu Zhou et al. utilized CSI features to virtually tune the coverage shape into disk-like [
27]. Speed Independent Entity Detection (SIED) extracted a novel feature from the whole wireless channel and transformed human detection into a probabilistic problem to achieve a high detection accuracy [
5]. AR-Alarm utilized a self-adaptive learning mechanism to achieve intrusion detection without the need of re-calibration [
13].
Besides human detection, wireless signals can be used in indoor localization, activity recognition and even human identification. Abdel-Nasser et al. utilized CSI to provide a localization approach with a high accuracy leveraging only a single pair of transceiver [
28]. CSI-MIMO utilized frequency diversity of CSI to construct the fingerprint of different locations and achieved a localization accuracy of 0.95 m [
29]. SpotFi computed the Angle of Arrival (AoA) of multipath components of different antennas and improved the localization accuracy to 40 cm [
30]. HiDFPL proposed a measurement to represent the sensitivity of the receiver and enhanced the localization accuracy [
31]. Xuyu Wang et al. proposed PhaseFi, a fingerprinting system, using phase information of CSIs and incorporated a greedy algorithm to train the weights of a deep network [
32]. Rui Zhou et al. proposed an indoor localization system based on CSI and SVM [
33]. Density-based Spatial Clustering Of Applications With Noise (DBSCAN) was utilized in the system to reduce the noise in CSIs.
CSI based human Activity Recognition and Monitoring (CARM) was proposed based on CSIs of wireless channel that quantified the relationship between the movement speeds of different body parts and activities, and it had the ability to recognize human activities [
34]. Activity recognition has a wide range of applications, such as somatosensory games. Wi-Play extracted CSI waveforms from commercial WiFi devices to model some specified activity and achieved an activity recognition system [
35]. Wifi-based GEsture Recognition (WiGeR) utilized the fluctuation scheme of CSIs generated by the moving of human hands to recognize gestures [
36]. Smokey leveraged WiFi signals and had the ability to recognize smoking activity even in the non-line-of-sight (NLOS) and through-wall environments [
37]. Wi-Chase utilized the CSIs from all subcarriers to achieve a higher activity recognition accuracy [
38].
It is confirmed that human’s gait is unique among different people, thus it can be used to identify the human’s identity. WifiU was presented to construct the gait profiles of different people utilizing the unique variations in the CSIs [
39]. WiWho was presented as a framework of human identification utilizing human’s gait extracted from CSIs [
40]. FreeSense combined Principal Component Analysis (PCA), Discrete Wavelet Transform (DWT), and Dynamic Time Warping (DTW) to achieve a nine-user human identification [
41]. Wii extracted time and frequency-domain features and used time−frequency analysis to achieve an accurate human identification system [
7].
Although there have been quantities of work on human detection, they only perform well when the people move in regular patterns. When an intruder appears, he is more likely to move in an irregular way. As a result, a more robust human detection system is proposed in this paper to meet the challenges of intruder detection.
3. Preliminary
CSI is leveraged in this study, and we will give a brief introduction of the background knowledge in this section.
The wireless signals propagate through multiple paths from the transmitter to the receiver in a typical indoor scenario. As a result, the received signal is the superposition of the signals from LOS path and several reflection paths. OFDM framework is the basis of 802.11 n wireless networks, in which our system works. In this framework, the wireless channel can be descripted by a Channel Impulse Response (CIR) in the time domain. Under the assumption of time-invariant, CIR can be expressed as:
where
,
, and
denote the amplitude, phase and time delay of the signal from
path, respectively;
is the total number of paths;
is complex Gaussian white noise; and
is the Dirac delta function.
Nevertheless, precise CIR can be extracted only from dedicated devices rather than commodity infrastructures. To overcome this limitation, Channel Frequency Response (CFR) can be extracted from frequency domain, which can model the wireless channel. CFR contains amplitude−frequency response and phase−frequency response. Under the assumption of infinite bandwidth, CIR is equivalent to CFR, and CFR can be transformed by Fast Fourier Transform (FFT) from CIR: [
20]
We can obtain CFRs in the format of CSI:
where
N is the number of subcarriers in the wireless network.
The CSI is composed of amplitude and phase of a subcarrier:
where
is the central frequency of the subcarrier, and
represents its phase. Thus, a group of CSIs,
, denote
K sampled CFRs in subcarrier level.
4. System Design
4.1. System Overview
The framework of RDFID is presented in
Figure 2. The system has four modules: pre-processing; feature extraction; classification; and post-processing. There are various kinds of noise in the raw collected CSI data, and most noise is removed in pre-processing module. We extract wavelet variance as the real-time feature from frequency domain in feature extraction module. In the classification module, a portion of data is utilized to train a system to be universal that can be adaptive to different scenarios. In the post-processing module, the classification result is further processed to be closer to reality.
The system can work in typical indoor scenarios with only one pair of commodity WiFi devices, which include a wireless router and a laptop. The wireless router is the Transmit Xmt (TX) that supports Institute of Electrical and Electronic Engineers (IEEE) 802.11n protocol, while the laptop is the Receive Xmt (RX) that is equipped with Intel 5300 network interface card (NIC). The WiFi devices keep transmitting data to collect CSIs in the monitoring area, and the system estimated intruder existence according to the extracted feature.
4.2. Pre-Processing
The CSI data is extracted from the respond packets of Internet Control Messages Protocol (ICMP) packets. As a result, the number of the group of CSIs is the same as that of ICMP packets theoretically. However, during data collection period, we find that the number of collected CSI records is larger than that of transmitted ICMP packets we had set in advance. In order to calibrate the frequency of the collected data, we conduct the linear interpolation in the raw data and it has a unified frequency. In 802.11 n wireless networks, there are several subcarriers transmitting signals at the same time under the OFDM framework. The subcarriers are independent theoretically. However, the CSIs of adjacent subcarriers have some relationships. In consequence, PCA is used to extract independent data. The related CSI streams can be combined into several independent principle components. For each ICMP packet, a matrix of 3 × 30 constructed by CSIs can be extracted from the firmware. It can be further reshaped into a 1 × 90 vector. For a certain time window, n ICMP packets have been received, and we can obtain an n × 90 matrix. During the evaluation of the principle components, we find that in most cases the first principle component can give an 80% contribution rate. As a result, we use the first principle component as the representative data.
Unfortunately, there still exist some kinds of noises in the first principle component, and they have negative impact on detection rate. The one that has the most significant impact is high frequency noise induced by environment changes other than human movement. The movement of torso, arms, and legs cause most of signal reflections. The frequency of the movements is lower than 10 Hz according to our observation. As a result, a low pass filter is utilized to filter out the high frequency noise from the collected data with the frequency higher than 10 Hz.
4.3. Feature Extraction
A proper feature is critical in classification tasks. Generally, the moving speed of a person is constant in a short period, and some periodicity exists when the person is moving. For instance, when the person walks, two steps construct a period. However, it is a challenging task to analyze the periodicity directly from the waveform of the wireless signals. During our early exploration, we find that besides time-domain features, frequency-domain features can better characterize the waveforms in intrusion detection. As a result, in order to explore a scenario independent feature, we utilize time−frequency analysis on the waveform. Continuous Wavelet Transform (CWT) combined with wavelet variance is a proper tool to analyze the periodicity of the waveform. First, the wavelet coefficient of the first principle component of the CSIs after low-pass filter (
cpl) is calculated utilizing CWT in Equation (5):
where
is the first principle component of the CSIs after low-pass filter (
cpl),
a and
b are scale and time, respectively.
is the wavelet function, and db6 (Daubechies) wavelet [
42] is selected as it provides the best performance after we have tried different wavelet functions.
As shown in
Figure 3, it can be clearly seen that some periodicity exists in the waveform after we conduct continuous wavelet transform. However, it is necessary to quantitatively calculate the significance of the periodicity to confirm that the periodicity is caused by human behaviors.
Wavelet variance is widely used in meteorology to calculate the periodicity of precipitation. It reflects the distribution of the power of the wavelet coefficients of various scales. As a result, it can also describe the significance of the periodicity of human motion. The wavelet variance is calculated as Equation (6):
where
is the power of the wavelet coefficient of scale
a at time
b.
During our experiment, we find that the distribution of the wavelet variance is different among whether there is human motion as shown in
Figure 4. In consequence, the wavelet variance is a proper feature for intrusion detection.
4.4. Training and Classification
As the distribution of the wavelet variance when there is human motion is different from that of static scenario, the Gaussian Mixture Model (GMM) is an appropriate classifier. In this GMM, there are two Gaussian models, one is static model and the other is human motion model. The moving data of different volunteers in different moving patterns and the data collected in the static scenario construct the training data. The GMM only need to be trained once, and it can be used in different scenarios without being re-trained. As a result, after a trained GMM is generated, the intrusion detection system is unsupervised. In the training phase, the vectors of wavelet variance of different scales and the ground truth are utilized to train the GMM. In the classification phase, the inputs are only the vectors of wavelet variance, while the outputs are the detection results whether there exists human motion.
In the end of classification, a post-processing procedure is added to improve the detection accuracy. In this procedure, it is assumed that a person cannot appear and disappear suddenly. As a result, an additional window beyond the detection window is utilized to reduce the detection mistakes. For example, 0 and 1 represent static and intrusion, respectively. If the detection result is 11011 in this additional window, we can consider there always exists intrusion in this window. The cost of this procedure is the time delay in detection, but the detection accuracy can be higher.
6. Discussion
We did a set of evaluations in this work and demonstrated the effectiveness of RDFID to detect human motion of different moving patterns using WiFi signals. However, there are still some limitations in RDFID. In this section, we will give a discussion about the limitations and potentials of RDFID.
Although the approach can achieve a high intrusion detection accuracy, it may be influenced by several factors.
First, the relative location of the intruder and transceivers can affect the detection accuracy. There exists a relationship between the impact of the intruder to the signal transmission and the distance of the intruder to the transceivers. When the intruder moves far away from the transceivers or the first Fresnel zone, it becomes more difficult to extract effective features from the collected CSI of the ambient wireless signal. As a result, the detection accuracy suffers a degradation when the distance of the intruder to the transceivers.
In addition, in real scenarios there may exist more than one intruder. Nevertheless, the movement of multiple intruders will break the periodicity of the received CSI. In consequence, the detection performance will be affected directly.
Despite these limitations, WiFi signal-based intrusion detection systems have much potential in a smart home. In our future work, we will explore more effective features that less affected as much by the distance of the intruder to the transceivers and the number of the intruders in the environment to make the approach more robust in smart home applications.