Table 2:Variance of different features extracted from keystrokes of keys a-z collected from user 10 eatures Mean amplitude 2n4 0018 460440 2 0003000i n0omoi .000四 001g6 4.52319192.40210 93159 28 Zero Crossings 3196 1299433662r393 314 125 1525 92 122 177 each of the 37 keys(26 alphabets,10 digits and 1 space bar) We also observe from this figure that the kevstrokes that are by pressing that key multiple times.After this,these users missed are usually those for which fingers move very little typed the sentence S="the quick brown fox jumped over when typing.For example,in pressing keys 'a','d',f','i',j the lazy dog"two times,without spaces. and 'x'the hands and fingers move very little,and thus the To evaluate how the number of training samples impact variations in the CSI values sometimes go undetected.Fig- the accuracy,we collected 80 samples for each of the 37 keys ure 6(b)shows the keystroke extraction rate for each user from User 10.Afterwards,this user typed each of the follow- averaged over all 37 keys.The experimental results show ing sentences 5 times,without spaces:S ="the quick brown that our keystroke extraction algorithm is robust because it fox jumps over the lazy dog",S2="nobody knew why the consistently achieves high performance over different users candles blew out",S3="the autumn leaves look like golden without requiring any user specific tuning of system para- snow",S4 ="nothing is as profound as the imagination" meters. and Ss ="my small pet mouse escaped from his cage".We asked users to type naturally with multiple fingers but only 8.4 Classification Accuracy press one key at a time while keeping the average keystroke We evaluate the classification accuracy of Wikey through inter-arrival time at 1 second.After recording the CSI time two sets of experiments.In the first set of experiments,we series for each of the above experiments,we first applied our build classifiers for each of the 10 users using 30 samples keystroke extraction algorithm on those recorded CSI time and measure the 10-fold cross validation accuracy of those series to extract the CSI waveforms for individual keys and classifiers.In the second set of experiments,we build clas- then extracted the DWT based shape features from each of sifier for user 10 while increasing the number of samples the extracted keystroke waveforms. from 30 to 80 in order to observe the impact of increase in 8.3 Keystroke Extraction Accuracy the number of training samples on the classification accur- acy.Cross validation automatically picks a part of data for We evaluate the accuracy of our keystroke extraction al- training and remaining for testing and does not use any data gorithm in terms of the detection ratio,which is defined as in testing that was used in training.Recall that the WiKey the total number of correctly detected keystrokes in a CSI uses kNN classifiers for recognizing keys.In all of following time series divided by the total number of actual keystrokes. experiments,we set =15. The detection ratio of our proposed algorithm is more than 97.5%.Figure 6(a)shows the color map showing the percent- 8.4.1 Accuracy with 30 Samples per Key age of the missed keystrokes of all 37 keys for all 10 users. We evaluate the classification accuracy of WiKey in terms of average accuracy per key and average accuracy on all keys of any given user.We also present confusion matrices resulting from our experiments.A confusion matrix tells us which key was recognized by WiKey as which key with what percentage.We calculate the average accuracy per key by taking the average of confusion matrices obtained from all users and average accuracy on all keys of any given user by averaging the accuracy on all keys within the confusion (a)Colormap for missed keys (b)Keystroke extraction rates matrix of that user.For each user,we trained each classifier per user averaged over all keys using features from 30 samples of each key.We conducted Figure 6:Keystroke extraction results our experiments on all 37 keys as well as on only 26 alphabet keys and performed 10-fold cross validation to obtain the The darker areas represent higher rate of missed key- confusion matrices. strokes.We can observe from this figure that the number Wikey achieves an overall keystroke recognition accuracy of missed keystrokes vary for different individuals depend- of 82.87%in case of 37 keys and 83.46%in case of 26 al- ing upon their typing behaviors.For example we observed phabetic keys when averaged over all keys and users.Fig- that the keystrokes of user 4 were missed in higher per- ure 7 shows the recognition accuracy for each key across all centage with average detection ratio of 91.8%whereas the users for the 26 alphabetic keys.Similarly,Figure 8 shows keystrokes of user 10 were not missed at all with average de- the recognition accuracy for each key across all users for all tection ratio of 100%calculated over all 37 keys.The lower 37 keys.Figure 9 shows the average recognition accuracy extraction accuracy for user 4 shows that more keystrokes achieved by each user for both 26 keys and 37 keys.We ob- were missed.which is due to the significant difference in his serve that the recognition accuracy for 26 alphabetic keys is typing behavior compared to other users.The accuracy of on average greater than the recognition accuracy for the all our scheme for such a user can be increased significantly by 37 keys.This is because the keystroke waveforms of the digit tuning the parameters of our algorithm for the given user keys(0-9)often show similarity with keystroke waveforms ofTable 2: Variance of different 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 0.00029 4E-04 0.0003 1E-04 4E-04 0.0002 0.0003 8E-04 5E-04 5E-04 5E-04 2E-04 5E-04 3E-04 3E-04 5E-04 0.0003 0.00018 6E-04 4E-04 3E-04 1E-04 3E-04 4E-04 6E-04 4E-04 Second central moment 0.00513 0.003 0.0011 0.001 0.007 0.0028 0.1008 0.012 0.005 0.009 0.016 0.006 0.002 0.003 0.002 0.007 0.0017 0.00041 0.03 0.003 0.001 0.006 0.002 0.007 0.003 0.005 Third central moment 0.00155 9E-04 0.0001 2E-04 0.002 0.0033 0.7021 0.002 0.001 0.007 0.009 0.017 5E-04 3E-04 5E-04 0.003 0.0003 7.70E-05 0.024 0.003 1E-04 0.015 9E-04 0.001 6E-04 0.003 RMS deviation 0.0108 0.006 0.0031 0.004 0.011 0.0038 0.0348 0.011 0.011 0.01 0.012 0.009 0.006 0.005 0.004 0.008 0.0042 0.00196 0.026 0.004 0.004 0.008 0.004 0.007 0.007 0.007 Energy 3874.59 2283 816.91 912.4 5204 2160 76863 9315 3925 6846 12094 4883 1679 2153 1150 5048 1296.2 308.95 23201 2181 886.9 4714 1403 5166 2100 4147 Entropy 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 Zero Crossings 26.3859 12.36 33.196 12.94 9.433 6.2627 3.9943 3.585 44.91 13.14 12.58 31.14 28.51 15.25 29.24 24.09 21.673 17.3847 12.21 17.7 36.85 21.63 27.71 13.67 31.49 6.529 each of the 37 keys (26 alphabets, 10 digits and 1 space bar) by pressing that key multiple times. After this, these users typed the sentence S1 = “the quick brown fox jumped over the lazy dog” two times, without spaces. To evaluate how the number of training samples impact the accuracy, we collected 80 samples for each of the 37 keys from User 10. Afterwards, this user typed each of the follow￾ing sentences 5 times, without spaces: S1 =“the quick brown fox jumps over the lazy dog”, S2 = “nobody knew why the candles blew out”, S3 = “the autumn leaves look like golden snow”, S4 = “nothing is as profound as the imagination” and S5 = “my small pet mouse escaped from his cage”. We asked users to type naturally with multiple fingers but only press one key at a time while keeping the average keystroke inter-arrival time at 1 second. After recording the CSI time series for each of the above experiments, we first applied our keystroke extraction algorithm on those recorded CSI time series to extract the CSI waveforms for individual keys and then extracted the DWT based shape features from each of the extracted keystroke waveforms. 8.3 Keystroke Extraction Accuracy We evaluate the accuracy of our keystroke extraction al￾gorithm in terms of the detection ratio, which is defined as the total number of correctly detected keystrokes in a CSI time series divided by the total number of actual keystrokes. The detection ratio of our proposed algorithm is more than 97.5%. Figure 6(a) shows the color map showing the percent￾age of the missed keystrokes of all 37 keys for all 10 users. Users Keys SPa b c d e f g h i j k l mn o p q r s t u v w x y z 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 (a) Colormap for missed keys 1 2 3 4 5 6 7 8 9 10 86 88 90 92 94 96 98 Users Extraction rate (%) (b) Keystroke extraction rates per user averaged over all keys Figure 6: Keystroke extraction results The darker areas represent higher rate of missed key￾strokes. We can observe from this figure that the number of missed keystrokes vary for different individuals depend￾ing upon their typing behaviors. For example we observed that the keystrokes of user 4 were missed in higher per￾centage with average detection ratio of 91.8% whereas the keystrokes of user 10 were not missed at all with average de￾tection ratio of 100% calculated over all 37 keys. The lower extraction accuracy for user 4 shows that more keystrokes were missed, which is due to the significant difference in his typing behavior compared to other users. The accuracy of our scheme for such a user can be increased significantly by tuning the parameters of our algorithm for the given user. We also observe from this figure that the keystrokes that are missed are usually those for which fingers move very little when typing. For example, in pressing keys ‘a’, ‘d’, ‘f’, ‘i’, ‘j’ and ‘x’ the hands and fingers move very little, and thus the variations in the CSI values sometimes go undetected. Fig￾ure 6(b) shows the keystroke extraction rate for each user averaged over all 37 keys. The experimental results show that our keystroke extraction algorithm is robust because it consistently achieves high performance over different users without requiring any user specific tuning of system para￾meters. 8.4 Classification Accuracy We evaluate the classification accuracy of WiKey through two sets of experiments. In the first set of experiments, we build classifiers for each of the 10 users using 30 samples and measure the 10-fold cross validation accuracy of those classifiers. In the second set of experiments, we build clas￾sifier for user 10 while increasing the number of samples from 30 to 80 in order to observe the impact of increase in the number of training samples on the classification accur￾acy. Cross validation automatically picks a part of data for training and remaining for testing and does not use any data in testing that was used in training. Recall that the WiKey uses kNN classifiers for recognizing keys. In all of following experiments, we set k = 15. 8.4.1 Accuracy with 30 Samples per Key We evaluate the classification accuracy of WiKey in terms of average accuracy per key and average accuracy on all keys of any given user. We also present confusion matrices resulting from our experiments. A confusion matrix tells us which key was recognized by WiKey as which key with what percentage. We calculate the average accuracy per key by taking the average of confusion matrices obtained from all users and average accuracy on all keys of any given user by averaging the accuracy on all keys within the confusion matrix of that user. For each user, we trained each classifier using features from 30 samples of each key. We conducted our experiments on all 37 keys as well as on only 26 alphabet keys and performed 10-fold cross validation to obtain the confusion matrices. WiKey achieves an overall keystroke recognition accuracy of 82.87% in case of 37 keys and 83.46% in case of 26 al￾phabetic keys when averaged over all keys and users. Fig￾ure 7 shows the recognition accuracy for each key across all users for the 26 alphabetic keys. Similarly, Figure 8 shows the recognition accuracy for each key across all users for all 37 keys. Figure 9 shows the average recognition accuracy achieved by each user for both 26 keys and 37 keys. We ob￾serve that the recognition accuracy for 26 alphabetic keys is on average greater than the recognition accuracy for the all 37 keys. This is because the keystroke waveforms of the digit keys (0-9) often show similarity with keystroke waveforms of