DTW Accutac o BTW nn time (scaled) 8：1141822283如 13.5070010101531019的21029 14 Ner Nei时cour k Down-3tpn3Leh作 Neares1 Neighbour k Fean Eec Fig.8:Accuracy changes with Fig.9:Accuracy and scaled run-Fig.10:Running time compari-Fig.11:The Fl-score of certain sampling lengths(nEp)and num-ning time using DTW change with son between our scheme and linear category of features is excluded. bers of nearest neighbours. sampling lengths (nDrw ) scanning. (a)TPR v.s.Training Size (b)FPR v.s.Training Size (a)FI-Scores for Gesture Nod (b)FI-Scores for Gestare Shake 00% 4% 099 75% 03 097 Average-TPR-OCSVM Average-TPR-OCESVN 50% 2 25% 01% 093 0.10.20304050.60.70.8091.0 0 03 0.10.2030.4050.60.7080.91.0 Fig.12:The average TPR and FPR change with different ratios of Fig.13:F1-Scores of one-class SVM with or without feature selection training samples. for gesture nod and shake. are small.Therefore,in our system,we build the training set if using Nod and Shake gestures.We are careful to bother passively and continuously in the background every time the no authorized user with occasional false positives.However, user performs those gestures.We employ OCESVM when since gestures are very short,cost nothing,and are easy to the size of training samples is insufficient,and fall back perform,we assume that the user is willing to go through to OCSVM for system overhead concern when the gathered authentication multiple times which can basically eliminate the training samples are adequate.This scheme eases the burden false positives.We compare authentication performance using of training on users significantly while maintaining high TPR two consecutive gestures with work [19]and [13],where both and low FPR at the same time. one class classifiers are used.Work [19]combines touchpad and voice commands to authenticate user in Google Glass.Our Single TPR FPR scheme requires fewer gestures,less effort,and produces better GlassGesture Nod 92.43%(+/-3.09)0.09%(+/-0.22 GlassGesture Shake 92.33%(+/-3.32)0.17%(+/-0.33) result.Work [13]is about touchscreen-based authentication on GlassGesture Left3 89.08%(+/-6.360.48%(+/-0.79 smartphone,while we show that we can achieve competitive GlassGesture Right3 89.61%(+/-5.99)0.52%(+1-0.87 performance using head gestures on Google Glass.Another Multiple and Comparison TPR FPR work [18]uses a two-class SVM classifier,which only reports GlassGesture 99.16% 0.61% 2 gesturesi the average error rate (AER,defined as (1-tpr+fpr))as Touchpad+ Voice (5 events)[19] 97.14% 1.27% 0.04 while using 5 touchpad gestures on Glass.Our scheme louchscreen 98.2% 1.1% requires fewer gestures,and better result when multiple ges- GEAT (3 gestures)[13] tures are combined. TABLE II:FPR and TPR of authentication on two gestures. Impact of Peak Features.With the intention to investigate the impact of peak features.we use an feature-excluding Authentication against Type-II attacker.In order to method to verify the effectiveness of peak features.We collect understand the authentication performance,we evaluate the number of true positive,false negative and false positive authentication against Type-II attackers,which are more pow- to calculate the Fl-score as a metric to show the overall erful than Type-I attackers.We utilize the full size of the data performance of the classification.In Fig.11,the Fl-score is set to train the model with 10-fold cross-validation for each lowered the most when peak features are excluded.Some other user.While training model for a certain user,we use data important features are mean,energy,kurtosis and skewness. samples from all other users as Type-II attacking trials.The Impact of Feature Selection.To show how our feature result is shown in Table II in metrics of TPR and FPR and selection method helps in our case,we compare Fl-scores of compared with several existing works.From the result,we classification with and without feature selection.From Fig. can tell that our authentication system can identify authorized 13,we can find that for majority of users (13/18 and 12/18 users with with a high TPR as average 92.38%and defend respectively),feature selection improves the classification. against Type-II attackers with a low FPR as average 0.13% Imitator Attack (Type-IID).In this evaluation,we want to