20 409 lemporal units (a)Keystroke waveforms for key i (b)Keystroke waveforms for key o (c)DWT features of i (d)DWT features of o from 24 waveforms from 2nd waveforms Figure 5:Feature extraction from 24 keystroke waveforms extracted from TX-1,RX-1 for I and O Table 1:Average values of features extracted from keystrokes of keys a-z collected from user 10 h 00 m 入 Mean amplitu 0024 nd central moment 03000100s302 0. hird central moment 002-00 000096-0010029 0.0609190.05001010.05 0.02 0.010.010.040.0060.0030.0980.101 0.0290.02 00.020.04 MS deviation027080278202s5050.4244079905060204720570.320.30.32☒0.290.4303137Q222Q4720.430.220.3350.3060.503045 Entropy97697629761697629769.76169766972976297@9.769.769.769.7629.769.7697616976297629729.7629.7629.7629.7696976 Zero Crossin869323622544554053420906013.7109.063.82.9855.635216.7515s1436.48101755 waveforms by performing optimal alignment between them. tenna pairs.To classify a detected keystroke,WiKey feeds Using DTW distance as the comparison metric between key- the shape features of that keystroke to their corresponding stroke shape features,WiKey trains an ensemble of k-nearest kNN classifiers and obtains a decision from each classifier in neighbour (kNN)classifiers using those features from all TX- the ensemble.Each kNN classifier searches for the majority RX antenna pairs.WiKey obtains decisions from each clas- class label among k nearest neighbors of the corresponding sifier in the ensemble and uses majority voting to obtain shape feature using DTW distance metric.WiKey calculates final result.Next,we first explain how we apply DTW on the final result through majority voting on the decisions of the keystroke shape features and then explain how we train all kNN classifiers in the ensemble. the ensemble of classifiers. 8.IMPLEMENTATION EVALUATION 7.1 Dynamic Time Warping DTW is a dynamic programming based solution for ob- 8.1 Hardware Setup taining minimum distance alignment between any two wave- We implemented our scheme using off-the-shelf hardware forms.DTW can handle waveforms of different lengths and devices.Specifically,we use a Lenovo X200 laptop with Intel allows a non-linear mapping of one waveform to another by Link 5300 WiFi NIC as the receiver that connects to the minimizing the distance between the two.In contrast to Eu- three antennas of the X200 laptop.The X200 laptop has clidean distance,DTW gives us intuitive distance between 2.26GHz Intel Core 2 Duo processor with 4GB of RAM and two waveforms by determining minimum distance warping Ubuntu 14.04 as its operating system.We used TP-Link TL- path between them even if they are distorted or shifted WR1043ND WiFi router as transmitter operating in 802.11n versions of each other.DTW distance is the Euclidean dis- AP mode at 2.4GHz.We collect the CSI values from the Intel tance of the optimal warping path between two waveforms 5300 NIC using a modified driver developed by Halperin et calculated under boundary conditions and local path con- al.[24.The transmitter has 2 antennas and the receiver has straints [25.In our experiments,DTW distance proves to 3 antennas,.元.e.,Mr=2 and MR=3.This gives3×2×3= be very effective metric for comparing two shape features of 18 classification models for each key in our evaluations. different keystrokes.WiKey uses the open source implement- We place the X200 laptop at a distance of 30 cm from ation of DTW in the Machine Learning Toolbox (MLT)by the keyboard such that the back side of its screen faces the Jang 26.WiKey uses local path constraints of 27,45,and keyboard on which the users type and its screen is within the 63 degrees while determining minimum cost warping path line-of-sight (LOS)of the WiFi router it is connected to.The between two waveforms.For the features extracted for keys distance of WiFi router from the target keyboard is 4 meters i'and 'o'shown in figures 5(a)and 5(b).the DTW distance The CSI values are measured on ICMP ping packets sent among features of key 'i'was 18.79 and the DTW distance from the WiFi router,i.e.,the TP-Link TL-WR1043ND.to among features of key 'o'was 19.44.However,the average the laptop at high data rate of about 2500 packets/s.Setting DTW distance between features of these keys was 44.2. a higher ping frequency leads to higher sampling rate of CSI. 7.2 Classifier Training which ensures that the time resolution of the CSI values is high enough for capturing maximum details of different type To maximize the advantage of having multiple shape fea- of keystrokes tures per keystroke obtained from multiple transmit-receive antenna pairs,we build separate classifiers for each of those 8.2 Data Collection shape features.We build an ensemble of 3 x Mr x MR clas- To evaluate the accuracy of Wikey,we collected train- sifiers using kNN classification scheme.WiKey requires the ing and testing dataset from 10 users.These 10 users were user to provide training data for the keystrokes to be recog- general university students who volunteered for the experi- nized and each classifier is trained using the corresponding ments and only 2 out of them had some know how of wire- features extracted from CSI time series from all TX-RX an- less communication.Users 1-9 first provided 30 samples for0 200 400 600 800 −2 −1 0 1 2 3 Sample Value 0 200 400 600 800 −2 −1 0 1 2 3 sample Value 1 2 3 1 2 3 (a) Keystroke waveforms for key i 0 200 400 600 800 −2 −1 0 1 2 3 sample Value 0 200 400 600 800 −4 −2 0 2 4 6 Sample Value 1 2 3 1 2 3 (b) Keystroke waveforms for key o 0 50 100 −4 −3 −2 −1 0 1 2 temporal units Approximation Coeefficients (c) DWT features of i from 2nd waveforms 0 50 100 −4 −2 0 2 4 temporal units Approximation Coefficients (d) DWT features of o from 2nd waveforms Figure 5: Feature extraction from 2 nd keystroke waveforms extracted from TX-1, RX-1 for I and O Table 1: Average values of features extracted from keystrokes of keys a-z collected from user 10 Features a b c d e f g h i j k l m n o p q r s t u v w x y z Mean amplitude -0 -0.04 0.0124 -0.03 0.045 -0.043 -0.076 -0.06 0.014 -0.03 0.03 -0.01 -0 0.032 0.02 0.03 -0.012 0.008 0.054 7E-04 -0.013 -0.02 -0 -0.1 -0.02 0.06 Second central moment 0.08 0.133 0.0801 0.083 0.156 0.1818 0.6523 0.263 0.12 0.231 0.33 0.11 0.1 0.108 0.09 0.19 0.1022 0.051 0.245 0.192 0.062 0.12 0.097 0.26 0.09 0.21 Third central moment 0.02 -0.03 0.0036 -0.01 0.029 -0.06 -0.919 -0.05 -0.01 -0.1 0.05 0.02 -0 0.01 0.01 0.04 -0.006 0.003 0.098 -0.101 -0.01 0.029 0.023 -0 -0.02 0.04 RMS deviation 0.27 0.359 0.2782 0.285 0.385 0.4244 0.7899 0.506 0.332 0.472 0.57 0.32 0.3 0.323 0.29 0.43 0.3137 0.222 0.472 0.434 0.242 0.335 0.306 0.5 0.3 0.45 Energy 71.5 116.6 69.788 73.34 137.5 159.43 570.8 232.1 104.8 201.4 288 95.2 83.7 94.98 75.6 167 88.928 44.22 215.5 167.1 54.56 104.4 84.48 227 81.5 182 Entropy 9.76 9.762 9.7616 9.762 9.762 9.7616 9.7616 9.762 9.762 9.762 9.76 9.76 9.76 9.762 9.76 9.76 9.7616 9.762 9.762 9.762 9.762 9.762 9.762 9.76 9.76 9.76 Zero Crossings 11.8 6.913 12.363 6.225 6.4 4.375 4.075 3.4 12.08 9.088 6.05 13.7 10 9.063 13.8 12.9 11.85 15.41 6.35 12.85 16.75 11.88 14.3 6.48 10.1 7.55 waveforms by performing optimal alignment between them. Using DTW distance as the comparison metric between key￾stroke shape features, WiKey trains an ensemble of k-nearest neighbour (kNN) classifiers using those features from all TX￾RX antenna pairs. WiKey obtains decisions from each clas￾sifier in the ensemble and uses majority voting to obtain final result. Next, we first explain how we apply DTW on the keystroke shape features and then explain how we train the ensemble of classifiers. 7.1 Dynamic Time Warping DTW is a dynamic programming based solution for ob￾taining minimum distance alignment between any two wave￾forms. DTW can handle waveforms of different lengths and allows a non-linear mapping of one waveform to another by minimizing the distance between the two. In contrast to Eu￾clidean distance, DTW gives us intuitive distance between two waveforms by determining minimum distance warping path between them even if they are distorted or shifted versions of each other. DTW distance is the Euclidean dis￾tance of the optimal warping path between two waveforms calculated under boundary conditions and local path con￾straints [25]. In our experiments, DTW distance proves to be very effective metric for comparing two shape features of different keystrokes. WiKey uses the open source implement￾ation of DTW in the Machine Learning Toolbox (MLT) by Jang [26]. WiKey uses local path constraints of 27, 45, and 63 degrees while determining minimum cost warping path between two waveforms. For the features extracted for keys ‘i’ and ‘o’ shown in figures 5(a) and 5(b), the DTW distance among features of key ‘i’ was 18.79 and the DTW distance among features of key ‘o’ was 19.44. However, the average DTW distance between features of these keys was 44.2. 7.2 Classifier Training To maximize the advantage of having multiple shape fea￾tures per keystroke obtained from multiple transmit-receive antenna pairs, we build separate classifiers for each of those shape features. We build an ensemble of 3 × MT × MR clas￾sifiers using kNN classification scheme. WiKey requires the user to provide training data for the keystrokes to be recog￾nized and each classifier is trained using the corresponding features extracted from CSI time series from all TX-RX an￾tenna pairs. To classify a detected keystroke, WiKey feeds the shape features of that keystroke to their corresponding kNN classifiers and obtains a decision from each classifier in the ensemble. Each kNN classifier searches for the majority class label among k nearest neighbors of the corresponding shape feature using DTW distance metric. WiKey calculates the final result through majority voting on the decisions of all kNN classifiers in the ensemble. 8. IMPLEMENTATION & EVALUATION 8.1 Hardware Setup We implemented our scheme using off-the-shelf hardware devices. Specifically, we use a Lenovo X200 laptop with Intel Link 5300 WiFi NIC as the receiver that connects to the three antennas of the X200 laptop. The X200 laptop has 2.26GHz Intel Core 2 Duo processor with 4GB of RAM and Ubuntu 14.04 as its operating system. We used TP-Link TL￾WR1043ND WiFi router as transmitter operating in 802.11n AP mode at 2.4GHz. We collect the CSI values from the Intel 5300 NIC using a modified driver developed by Halperin et al. [24]. The transmitter has 2 antennas and the receiver has 3 antennas, i.e., MT = 2 and MR = 3. This gives 3×2×3 = 18 classification models for each key in our evaluations. We place the X200 laptop at a distance of 30 cm from the keyboard such that the back side of its screen faces the keyboard on which the users type and its screen is within the line-of-sight (LOS) of the WiFi router it is connected to. The distance of WiFi router from the target keyboard is 4 meters. The CSI values are measured on ICMP ping packets sent from the WiFi router, i.e., the TP-Link TL-WR1043ND, to the laptop at high data rate of about 2500 packets/s. Setting a higher ping frequency leads to higher sampling rate of CSI, which ensures that the time resolution of the CSI values is high enough for capturing maximum details of different type of keystrokes. 8.2 Data Collection To evaluate the accuracy of WiKey, we collected train￾ing and testing dataset from 10 users. These 10 users were general university students who volunteered for the experi￾ments and only 2 out of them had some know how of wire￾less communication. Users 1–9 first provided 30 samples for