Keywords

1 Introduction

With the wide deployment of wireless local area network (WLAN) and general support of WLAN protocol by various intelligent terminals, the intrusion detection with respect to the indoor target can be realized by using the existing WLAN infrastructure. Among the existing anonymous target intrusion detection techniques, the wireless local area network (WLAN) indoor target intrusion detection system [1,2,3,4] proposed by the University of Maryland performs outstandingly because it can effectively protect the user’s location privacy and work stably under non-line-of-sight and without special hardware at the same time. However, the main problem with this kind of algorithms is that the construction of the prior passive radio map takes a lot of manpower and time, which is a major barrier of WLAN intrusion detection systems deployment. On this basis, the WLAN indoor target intrusion detection algorithm proposed in this paper uses the adaptive-depth ray tree-based quasi-3D ray-tracing model to construct the passive radio map automatically, which requires less labor overhead compared with the traditional RSS feature database construction method. In addition, six signal characteristics of the passive radio map are constructed, which results in the better pattern recognition ability and learning convergence. The rest of this paper is structured as follows. In Sect. 2, we describe the proposed indoor WLAN intrusion detection method in detail, and the related experimental results are shown in Sect. 3. Finally, we conclude this paper in Sect. 4.

2 System Description

The overall flow of the system is shown in Fig. 1. First, a number of WLAN access points (APs) and monitor points (MPs) are arranged in the target area. Second, the GA algorithm is used to optimize the limited number of depth of the ray tree adaptively and the RSS characteristics under the indoor silence, and intrusion scenarios are constructed according to the optimized ray-tracing model. Then, the obtained RSS characteristics are used for probabilistic neural network (PNN) training. Finally, the trained PNN is used to classify the new observation RSS data by multiple classifications, so as to realize the intrusion detection and area localization.

Fig. 1
figure 1

Overall system flowchart

Full size image
Fig. 2
figure 2

Signal prediction flowchart

Full size image

2.1 Signal Prediction

Considering the limitations of the existing 2D and 3D ray-tracing models [5, 6] on the accuracy of the signal prediction and the complexity of the algorithm respectively, the quasi-3D ray-tracing model used in this paper first carries out the ray-tracing in the 2D projection plane, and then transforms it into the propagation path in the 3D space, and this process significantly improves the computational efficiency while guaranteeing the accuracy of prediction. In this case, as shown in Fig. 2, a quasi-3D ray-tracing model based on the adaptive-depth ray tree is proposed in this paper, considering two factors: the model accuracy and calculation efficiency.

The import of environmental information. Figure 3 gives a 3D modeling of a simple environment and the corresponding 2D projection results. The gray and black parts of the diagram represent the boundary face of environment and the indoor facilities, respectively. In addition, in order to ensure the integrity of the imported environmental information, the 3D vertex coordinates, height information and relative permittivity, conductivity and permeability of the corresponding material will be recorded.

Fig. 3
figure 3

2D projection from 3D modeling of environment

Full size image

Optimization of the limited number of depth. In order to significantly improve the computing efficiency of the ray tree, the GA algorithm is used to optimize the limited number of depth of the ray tree in different environments. Specifically, first, the limited number of depth is initialized to 1; secondly, all the vertical planes and vertical lines of the 3D modeling of the environment are numbered; besides, the number of the functional parts of each ray is spliced into a chromosome in chronological order, and the field strength of each ray that reaches the MP is used as the fitness of its corresponding chromosome; then, the contribution rate of the ray to the field strength at MP under the condition of the current limited number of depth is calculated by Algorithm 1; finally, determine whether the contribution rate of the ray under the current limited number of depth is greater than the preset threshold, and if so, add 1 to the limited number of depth and repeat the above steps, otherwise the current limited number of depth is the optimal limited number of depth (or the optimal ray order). In Algorithm 1, the fitness of each chromosome in the current population is calculated by the reverse ray-tracing method, and its calculation process is described in Algorithm 2.

figure a
figure b

Calculation of received signal power. In order to calculate the received signal power of MP, direct and non-direct rays are considered respectively. All the direct and non-direct rays within the n order are superimposed on the signal field strength, and the received signal power at MP can be obtained by the ray power summation method [8] as

$$\begin{aligned} {P_{total}} = {\sum \limits _{i = 1}^l {\left( {\frac{{\lambda \left| {{{\mathbf {E}}_i}} \right| }}{{4\pi \left| {{{\mathbf {E}}_0}} \right| }}} \right) } ^2}, \end{aligned}$$
(1)

in which \(\mathbf {E}_0\) is the arrival signal field strength at 1 m from AP, and \(\mathbf {E}_i\) is the arrival signal field strength of the ith ray, l is the total number of rays.

2.2 Intrusion Detection

In this paper, the kernel density estimation method based on Bayesian decision theory is applied to train the PNN feature data under the indoor silence and intrusion scenarios.Footnote 1 In particular, the kernel density function is used to estimate the conditional probability of different states, and then the state of the maximum posterior probability is used as the PNN output [9] according to the Bias decision theory. In order to ensure the stability of RSS characteristic data between each pair of AP and MP, this paper uses a sliding window function to segment the original RSS dataFootnote 2 and calculates the mean, variance, maximum, minimum, maximum, and middle value of each segment data. On the basis of these six signal characteristics, six PNN structures are trained respectively. Finally, according to the voting criterion, the indoor target detection and location are realized by the multiclassification decision of the newly acquired RSS data.

3 Experimental Result

3.1 Environmental Layout

Figure 4 shows an experimental environment, in which two APs (AP1 and AP2 with model D-Link DAP 2310) and three MPs (MP1, MP2, and MP3 with model SAMSUNG GT-S7568) are placed at 2 and 0.5 m high, respectively. At each MP, 5 min of RSS data from each AP are collected separately under the indoor silence and intrusion scenarios.

Fig. 4
figure 4

Structure of experimental environment

Full size image
Fig. 5
figure 5

Change of overall fitness

Full size image

3.2 GA Optimization Result

Figure 5 shows the change of the overall fitness of each generation of population under the conditions of different values of \(\rho _{th}\) when the GA was used to calculate the ray contribution rate. The overall fitness is defined as the ratio of m to the population size M, and m is the number of chromosomes whose fitness is greater than the threshold value \(T_f\) in the population. It can be seen that with the increasing of population algebra, the overall fitness is on the rise and tends to be the same when the population algebra reaches 30. In addition, Table 1 compares the average time overhead required by the 3D ray-tracing model [5], the traditional 2D ray-tracing model [6], and the proposed method for the ray modeling between each pair of AP and MP under the condition of the limited number of depth of 3. It can be seen from the table that this method performs obviously better than the methods used in the literature [5, 6] in terms of time overhead.

Table 1 Average time cost for ray modeling between each pair of AP and MP
Full size table

3.3 Signal Prediction Result

Figures 6 and 7 compare the cumulative density function (CDF) of RSS prediction errors by the proposed method and the ones in [5, 6] under the limited number of depth of 3, from which we can find that the proposed method performs better than the others.

Fig. 6
figure 6

CDF of errors for AP1

Full size image
Fig. 7
figure 7

CDF of errors for AP2

Full size image

4 Conclusion

In this paper, we propose the adaptive-depth ray tree model, which can be used to adaptively construct a passive radio map for indoor WLAN intrusion detection. For one thing, we use the genetic algorithm to enhance the traditional quasi-3D ray-tracing model to depict the RSS variation under the indoor silence and intrusion scenarios with low labor and time cost. For another, six common signal features are allied to ensure the stability of RSS data and robustness of passive radio map. In future, we will continue to investigate a more effective passive radio map construction method to accurately locate multiple targets in the anonymous indoor WLAN environment.