Abstract
Device-free localization technology aims to find a target by analyzing the signal strength difference between transmitter and receiver deployed in the target area in advance. Up to now, device-free localization technology has been applied to a wide range of applications and scenarios, such as intrusion detection, environment modeling, and activity recognition. However, some sensors remain at potential risk that signal strength values of sensors have been tampered, or even devices sensors are physically damaged, which leads to inaccurate location results or a whole system crash. To solve the abovementioned problems, we design a CNN-based attack defense method for device-free localization, which can discover falsified signal strength values and error-prone devices. Firstly, we simulate a partial sensor attack or dropout in the device-free localization scenario. Then, we transform the localization problem into an image classification problem and use the convolutional neural networks (CNN) technique for abnormal detection. The experiment result shows that our algorithm can maintain high localization accuracy even under most sensor compromised and disconnected circumstances.
1. Introduction
Since localization has been a hot research topic from the last decade, a great number of work has been developed. Especially for outdoor localization, related positioning techniques are mature such as Global Position System (GPS), Cell Identification (Cell-ID), and Direction or Angle of Arrival (AOA), which can locate the target in an accurate and fast manner. However, the current indoor positioning system cannot achieve the high-precision localization result due to the complex surrounding environment and dense signal interference. For example, the traditional GPS cannot be used for localization in a complex building. Since the GPS signal travels from the satellite, the long-distance leads to the lower-barely GPS signal. Besides, heterogeneous Internet of Things (IoT) with various communication standards (e.g., WiFi, ZigBee, and Bluetooth) coexist in the building, which will result in severe signal interference. Moreover, the transmission scenario has a high penetration loss when the signal passes through rooms and stairs, significantly influencing the data quality for localization. As usual, the target must be equipped with a mobile device for indoor localization, which can receive and transmit related signals to the corresponding sides for distance measurement. However, sometimes the target does not carry a mobile device. At this time, the device-free localization technique rises to the surface, which localizes people without any specific devices or is actively involved in the whole process. As shown in Figure 1, device-free localization consists of wireless sensor networks (WSNs), where a great number of sensors have been deployed in advance for signal collection and analysis [1, 2]. For example, in the application of intrusion detection, once a thief gets into the house, the received signal strength indicator (RSSI) of on-premise sensors will be changed. Subsequently, all the sensing data will be transmitted back to the server, which aggregates all the RSSI and utilizes some pretrained deep learning networks to find out the current position of the thief. Finally, the administrator will be alerted. Rather than intrusion detection, the device-free localization also have been applied to fall detection and remote monitoring of the elderly, occupancy detection for energy-efficient heating, ventilation, and air conditioning (HVAC), and lighting [3–6].
However, the different RSS values cannot directly reflect the position information of target. Furthermore, the surrounding noise interferes with the signal RSS, deviating the final judgment of location distinction. Hence, we need to transform RSS values into vectors and adopt some deep learning techniques which can exclude noise inference to detect abnormal features more accurately. A great number of machine learning-based work have been proposed to solve the device-free localization problem such as the compressive sensing method [7], radio tomographic imaging (RTI) method [8], artificial neural network (ANN) method [9], and convolutional neural network (CNN) method [10]. In the abovementioned techniques, the CNN-based device-free localization is the most attractive since the CNN can quickly extract essential features from a cluster of image pixels and remove the embedded noise, which is helpful for RSS signal modeling and localization confirmation.
Although CNN-based device-free localization can find the target in high precision, some sensors are attacked to the extent that their signal strength values are tampered with, or some sensors are physically damaged and do not work properly. This situation can lead to low accuracy and unreliability in the positioning results. Even worse, the whole system cannot work properly during this period and miss some critical information. To solve this problem, based on the pretraining concept, we propose a CNN-based attack defense for device-free localization. First, deep learning techniques are used to transform the localization problem into an image classification problem. In contrast to traditional algorithms, we design and simulate a partial sensor attack or dropout scenario and train the neural network accordingly. This allows the system to function normally and obtain reliable and relatively accurate localization information even if some sensors are disconnected or attacked. The experiments and analysis show that our algorithm is more resilient to attacks than previous work, and the localization accuracy in the case of partial sensor failure or attack can be guaranteed. The contribution of this paper is illustrated as follows:(1)We propose a CNN-based attack defense for device-free localization, which can defend against sensor dropout and compromised attacks(2)We utilize the random seed to perturb the training data and generate an anti-interference device-free localization model in the pretraining stage(3)The experiment results on a real-world dataset verify that our proposed scheme can still achieve the accurate localization result in case of potential attacks
1.1. Organization of This Paper
The rest of this paper is organized as follows. Section 2 illustrates the related works about device-free localization. In Section 3, we formulate the DFL problem as a classification problem and also devise the BE scheme. Section 4 presents the algorithm. Section 5 evaluates the performance of our proposal. Finally, this work is concluded in Section 6.
2. Related Work
The concept of device-free localization has been proposed by Youssef et al. [11], which analyzed dynamic RF signals changed by the object movement in the environment and correlate their locations. To achieve better localization performance, in a real environment, Moussa et al. [12] modelled the localization problem into a fingerprint identification problem, which was solved by a new algorithm based on maximum likelihood estimator (MLE) and improved the performance of initial device-free localization system. Wilson et al. [8] firstly used radio tomographic imaging to extract the attenuation of RF wavelength when penetrating physical objects. The proposed method is able to realize the RF attenuation for object localization. Zhang et al. [13] proposed a new device-free object tracking system called RASS, which transforms the tracking field into various triangle areas and multiple channels are used to mitigate noise interference among different nodes. Seifeldin et al. [14] presented Nuzzer device-free localization system, which utilized RSS variance to measure the number of entities in the target area and find the exact position of these objects. Based on real-environment data sensing from surroundings, Guo et al. [15] proposed the RSS model, which consists of small-scale Rayleigh enhancement part and large-scale exponential attenuation part. The experiment results show their proposed framework can solve the unpredictable device-free localization issue resulted from the multipath interference. Wang et al. [16] designed a novel BGA framework, which consists of shadowed links and the prior information in the estimation period. The proposed BGA framework shows a good performance that tracking error is greatly reduced in the device-free localization phase. Ciuonzo et al. [17] utilized sensors to measure a random signal resulted from the distance between transmitter and the receiver, and then sent to a fusion centre for device-free detection performance improvement. For decentralized device-free detection, Ciuonzo et al. [18] proposed a general version of Rao test arises from maximization of Rao decision statistics family, which can improve the previous performance of device-free localization. Sabek et al. [19] transformed multiagent device-free localization problem into an energy-minimization framework, which can maintain temporal and spatial smoothness and consistency. The proposed solution can map the localization problem into a binary graph-cut problem. Moreover, the noise and accuracy are reduced and enhanced by the clustering techniques, respectively. Xiao et al. [20] designed a CSI-based passive device-free localization system called Pilot, which makes use of the features of CSI to monitor the abnormal information. Especially, the three blocks design can increase the efficiency of passive device-free localization problem. Wu et al. [21] proposed FILA, which is a cross-layer method that can achieve the aim of high accurate indoor localization in WLANs, which includes two parts: CSI-based propagation model and CSI-based fingerprinting. The FILA is implemented in commodity 802.11 NICs and the experiment evaluates the correctness and feasibility of the proposed method. Wang et al. [9] proposed a deep-learning-based indoor fingerprint framework according to the CSI value. Besides, the offline training and online localization phases. The experiment result shows the proposed method can achieve good performance under signal propagation environments.
3. Problem Statement
3.1. Device-Free Localization Problem Illustration
As shown in Figure 1, wireless sensor nodes which compose of the monitoring area sense the signal strength from the target and figure out the current position. Note that all the sensors transmit and receive wireless signals in turn. At the beginning, there was no object in the system, so the RSS matrix derived from the RSS matrix is empty, which can be defined as . As we know the number of wireless sensor is , the RSS measurement can be computed as . If the target enters into the measurement area, the total RSS matrix from all the wireless sensors can be computed as . Note that the RSS value from the th sensor to the th sensor can be defined as , whose summation derives the . As we have mentioned, if the target is not in the target area, the will be 0. When the target appears in the detection area, the surrounding signal transmitted and received by the corresponding sensors will be greatly influenced. For example, the initial value of the 5th sensor is will be changed to . Since the feature of RSS value has strong connection with the target position, the object can be localized by the RSS measurements at the end of device side. In the following sections, we will talk more in detail about this transformation.
3.2. Problem Modelling
The target localization problem can be transformed into RSS matrix computation. From Figure 1, we firstly collect the signal strength from all the sensors under two circumstances (e.g., target nonexistence and target existence). Note that the significant components of the signal are extracted by the background elimination method. Note that the background elimination can make the features of matrices variation more obvious. The signal variation can be formulated as . After collecting all data from the sensors, the RSS matrix can be established as . We can observe from the , RSS matrices can reflect the features of each target position. Once we consider the RSS matrix as the image matrix, we can transform object localization issue into image identification problem. Especially the different hot spots derived from the matrices can directly show the variations of RSS value. The feature image obtained from the RSS matrices also owns the same patterns for target position recognition. Hence, we can utilize some traditional image process technique such as CNN and ResNet to figure out the exact position of the target.
3.3. Data Collection
For easy measurement, we divided the whole sensing area into many small sections, which can be regarded as a class. grids compose of classes for image identification problem. The training data for target localization can be constructed as , where means that the number of experiment times. Note that is the training dataset.
3.4. Potential Attack
In this venue, we assume a potential attacker would like to interfere sensing data collection or crack some normal devices for the entire system. To simulate the behaviours of the adversary, we replace some random RSS value with training data, which is transformed to . Note that means the RSS value after interference. Subsequently, the training set also will be changed to . Hence, in our proposed method, we should conduct device-free localization even in under attack scenario accurately.
4. Proposed Method
CNN is a powerful tool to process image and has been applied to many application fields such as facial recognition, documents analysis, and climate estimation. In our proposed framework, we also utilize the CNN to train a device-free localization model in the pretraining process including normal data for attack defense. The detailed process is illustrated as follows:(1)Firstly, all the wireless sensors send and receive the signals from the remaining sensors collaboratively. If no object enters in the monitoring area, the RSS received by each sensor will be a constant value . Once an object appears, the RSS will be changed into another value, which is obviously different from the and . After aggregating all the , we start to analyse the exact location of the target.(2)Since some useless noise is included in the received , the original signal may be interfered for further analysis. As shows in Figure 2, when a set of is presented in the image format, the background color is very obvious and even may affect the judgment of object localization features. Hence, we need to apply some image preprocessing technique to remove these unnecessary factors. Background elimination (BE) is a widely used approach to remove background element and reduce storage cost. Thus, we utilize the BE technique to handle our figures as , where can reflect the significant location features of the object in some degree. Since the whole monitoring area has been divided into some grids, the location dataset can be represented as . Then, we can apply the CNN technique to establishing the device-free localization model in the pretraining process. Two-dimensional kernels in the convolution layer will scan the whole space in the source image to extract features, which will be the input for the subsample layer. Subsampling can reduce the reliance of precise positioning within feature maps and construct a more accurate device-free localization model. The related formulation is , where and are the parameters in the convolutional layer, means the convolution operation, and refers to the rectified linear unite (ReLU) as , where is obtained by each layer. Besides, we also adopt the filter concatenation method to extract features of different data. The final classification performance can be improved by this filter. The previous mentioned subsampling is executed by max pooling and dropout can decrease the positives of some neural, which can avoid overfitting situation if possible. Note that the cross-entropy error function is adopted in our framework to update our model as , where , , and mean the model prediction, label value. and cross-entropy error, respectively. Finally, we get a device-free localization model in the pretraining process.(3)In our assumed scenario, an adversary tries to interfere the received RSS value at each sensor’s side or some sensors may be compromised. To simulate the behaviours of the adversary, we add some random data into the partial training. Hence, RSS value will be changed to and the corresponding dataset turns . Then we use the pretrained device-free detection model to recognize the abnormal data. After a few turns training, the previous localization model can be transformed into the attack defense device-free localization model.(4)When a real adversary tries to conduct some attacks, the attack defense device-free localization model can quickly and accurately exclude the adverbial interference and localize the object without error.
The detailed process of the method is illustrated in Figure 2.
5. Performance Evaluation
In this section, we evaluate the performance of our proposed CNN attack defense device-free localization model under a real-world dataset collected by the SPANLab of The University of Utah. The experiments were performed in the TensorFlow 1.15.0 open source software, which is on the Windows 10 operation system with a Tesla T4 GPU and 16 GB of memory.
5.1. Dataset Description
In the used dataset, wireless sensor network consists of Crossbow TelosB nodes, whose protocol and frequency band are 2.4 GHz and IEEE 802.15.4, respectively. Moreover, this node takes the role of base station for data collection and process. From Figure 3, we can see the monitoring area is divided into 36 grids consisting of 28 TelosB nodes, whose total area is 21 21. Besides, the distance between two adjacent nodes is three feet. In our experiment, when a target enters into the range, the total measurement time for RSSI value is 30. The obtained training datasets are 25 samples, the testing data are 5, and the total selected reference points are 36. Moreover, the following equation is used to evaluate the accuracy of all the models: . The detailed reference point distribution can be referred to Figure 3. Note that the noise is embedded into original dataset for attack defense device-free localization method during the pretraining process.
5.2. Attack Defense Parameter Setting
For our attack defense device-free localization, we need to obtain some optimized parameters (i.e., the number of filters, the number of convolutional layers, subsampling layers, and kernel size) for CNN training. According to the expected result from the attack defense localization method, we can get the optimal parameters for our method, which is 2 convolutional layers, 99 concatenated convolutional filter size, 32 filter number for each layer, 100 epoches, learning rate, 300 batch size, and 0.4 dropout rate for avoiding overfitting.
5.3. Method Performance
Our proposed attack defense device-free localization protocol is trained in the pretraining process, where some random noise are replaced with normal RSS value. To evaluate the performance between our proposed CNN-based attack defense device-free localization method and BE-CNN, we need to adjust the signal-noise-rate for comparison. Firstly, we evaluate the converge rate for our proposed method and BE-CNN [22] under different noise distribution, where the related experiments are conducted on fix number of sensors. Moreover, we set the distance for two adjacent methods as 12 feet. As can be seen from Figure 4, when there is no noise interference in the environment, both our proposed method and BE-CNN can achieve 100% localization accuracy rate. When the noise level is increased, our proposed method can still achieve the accuracy which is still close to 100%. However, the localization accuracy for BE-CNN is suddenly decreased when the SNR is adjusted to a low value. When the SNR arrives the 5db, the BE-CNN can only achieve 90% accuracy rate compared with the nearly 100% that our proposed attack defense device-free localization method has achieved. Moreover, we also conducted the experiment to test the cumulative distribution function (CDF) performance for BE-CNN and our proposed attack defense device-free localization method under 5 db. As shown in Figure 5, after 30 times experiment, our proposed method can always achieve 100% localization accuracy, which is better than the performance achieved by BE-CNN in general. Finally, to evaluate the necessities and superiority for our selected CNN-based attack defense for device-free localization, we also change the deep learning method to SVM and KNN. From Figure 6, the result shows that our proposed CNN-based attack defense method can achieve the 100% localization accuracy, which is better than 89% and 51% achieved by the SVM and KNN, respectively. From the abovementioned multiple performance evaluation experiments, we can see our proposed CNN-based device-free localization protocol can defend sever noise interference and achieve the good localization accuracy.
6. Conclusion and Future Work
To defend potential device compromised and sever signal interference during device-free localization, we propose a CNN-based attack defense device-free localization method in this paper. The multiple experiments verify that our proposed method can achieve better localization accuracy compared with other deep learning-based technique and defend most of the possible attacks [23, 24]. In the future work, we will explore the popular CNN framework such as AlexNet or ResNet, which can maintain the localization accuracy under more sever circumstances.
Data Availability
The RSSI data used to support the findings of this study are included within the article
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work was supported by the National Natural Science Foundation of China under Grant no. 62001126 and in part by the China Postdoctoral Science Foundation Grant no. 2021M693617.