28th IFIP TC-11 SEC 2013 International Information Security and Privacy Conference, Jul 9, 2013
In this paper,we propose a novel gait authentication mecha-
nism by mining sensor resources on m... more In this paper,we propose a novel gait authentication mecha-
nism by mining sensor resources on mobile phone. Unlike previous works, both built-in accelerometer and magnetometer are used to handle mobile installation issues, including but not limited to disorientation, and misplacement errors. The authentication performance is improved by executing deep examination at pre-processing steps. A novel and effective segmentation algorithm is also provided to segment signal into separate
gait cycles with perfect accuracy. Subsequently, features are then extracted on both time and frequency domains. We aim to construct a light weight but high reliable model; hence feature subsets selection algorithms are applied to optimize the dimension of the feature vectors as well as the processing time of classification tasks. Afterward, the optimal
feature vector is classified using SVM with RBF kernel. Since there is no public dataset in this field to evaluate fairly the effectiveness of our mechanism, a realistic dataset containing the influence of mobile installation errors and foot gear is also constructed with the participation of 38 volunteers (28 males, 10 females). We achieved the accuracy approximately
94.93% under identification mode,the FMR, FNMR of 0%, 3.89% and processing time of less than 4 seconds under authentication mode.
With the tendency using machine learning on battery powered mobile devices for activity recogn... more With the tendency using machine learning on battery powered mobile devices for activity recognition (AR), recent achievements still remain restrict ions including low accuracy in single and multiple-subject AR and lacking of evidences about power consumption of feature extraction and classification. Thus, we contribute a novel method for extracting features in time domain and frequency domain. These features were then classified by Support Vector Machine (SVM) and dynamic time warping (DTW) method in order to find out the most effective combinations. Our own data and SCUTT-NAA dataset were used in our experiment. These results were also compared to previous achievements. Prototypes of the proposed methods were then implemented on a cell phone to measure power consumption. To reduce the energy overhead of continuous activity recognizing, we introduce an adaptive strategy by selecting an appropriate combination of flexible frequency and classification feature for each individual activity without decreasing its efficiency. We achieved an overall 28% of energy saving. Finally, to maintain these standards in cross-people AR, a combination of SVM classifier and K-means clustering algorithm is provided to personalize a known model . This method can enhance approximately 8% in accuracy compared with original cross-people AR.
28th IFIP TC-11 SEC 2013 International Information Security and Privacy Conference, Jul 9, 2013
In this paper,we propose a novel gait authentication mecha-
nism by mining sensor resources on m... more In this paper,we propose a novel gait authentication mecha-
nism by mining sensor resources on mobile phone. Unlike previous works, both built-in accelerometer and magnetometer are used to handle mobile installation issues, including but not limited to disorientation, and misplacement errors. The authentication performance is improved by executing deep examination at pre-processing steps. A novel and effective segmentation algorithm is also provided to segment signal into separate
gait cycles with perfect accuracy. Subsequently, features are then extracted on both time and frequency domains. We aim to construct a light weight but high reliable model; hence feature subsets selection algorithms are applied to optimize the dimension of the feature vectors as well as the processing time of classification tasks. Afterward, the optimal
feature vector is classified using SVM with RBF kernel. Since there is no public dataset in this field to evaluate fairly the effectiveness of our mechanism, a realistic dataset containing the influence of mobile installation errors and foot gear is also constructed with the participation of 38 volunteers (28 males, 10 females). We achieved the accuracy approximately
94.93% under identification mode,the FMR, FNMR of 0%, 3.89% and processing time of less than 4 seconds under authentication mode.
With the tendency using machine learning on battery powered mobile devices for activity recogn... more With the tendency using machine learning on battery powered mobile devices for activity recognition (AR), recent achievements still remain restrict ions including low accuracy in single and multiple-subject AR and lacking of evidences about power consumption of feature extraction and classification. Thus, we contribute a novel method for extracting features in time domain and frequency domain. These features were then classified by Support Vector Machine (SVM) and dynamic time warping (DTW) method in order to find out the most effective combinations. Our own data and SCUTT-NAA dataset were used in our experiment. These results were also compared to previous achievements. Prototypes of the proposed methods were then implemented on a cell phone to measure power consumption. To reduce the energy overhead of continuous activity recognizing, we introduce an adaptive strategy by selecting an appropriate combination of flexible frequency and classification feature for each individual activity without decreasing its efficiency. We achieved an overall 28% of energy saving. Finally, to maintain these standards in cross-people AR, a combination of SVM classifier and K-means clustering algorithm is provided to personalize a known model . This method can enhance approximately 8% in accuracy compared with original cross-people AR.
Uploads
Papers by Viet Vo
nism by mining sensor resources on mobile phone. Unlike previous works, both built-in accelerometer and magnetometer are used to handle mobile installation issues, including but not limited to disorientation, and misplacement errors. The authentication performance is improved by executing deep examination at pre-processing steps. A novel and effective segmentation algorithm is also provided to segment signal into separate
gait cycles with perfect accuracy. Subsequently, features are then extracted on both time and frequency domains. We aim to construct a light weight but high reliable model; hence feature subsets selection algorithms are applied to optimize the dimension of the feature vectors as well as the processing time of classification tasks. Afterward, the optimal
feature vector is classified using SVM with RBF kernel. Since there is no public dataset in this field to evaluate fairly the effectiveness of our mechanism, a realistic dataset containing the influence of mobile installation errors and foot gear is also constructed with the participation of 38 volunteers (28 males, 10 females). We achieved the accuracy approximately
94.93% under identification mode,the FMR, FNMR of 0%, 3.89% and processing time of less than 4 seconds under authentication mode.
nism by mining sensor resources on mobile phone. Unlike previous works, both built-in accelerometer and magnetometer are used to handle mobile installation issues, including but not limited to disorientation, and misplacement errors. The authentication performance is improved by executing deep examination at pre-processing steps. A novel and effective segmentation algorithm is also provided to segment signal into separate
gait cycles with perfect accuracy. Subsequently, features are then extracted on both time and frequency domains. We aim to construct a light weight but high reliable model; hence feature subsets selection algorithms are applied to optimize the dimension of the feature vectors as well as the processing time of classification tasks. Afterward, the optimal
feature vector is classified using SVM with RBF kernel. Since there is no public dataset in this field to evaluate fairly the effectiveness of our mechanism, a realistic dataset containing the influence of mobile installation errors and foot gear is also constructed with the participation of 38 volunteers (28 males, 10 females). We achieved the accuracy approximately
94.93% under identification mode,the FMR, FNMR of 0%, 3.89% and processing time of less than 4 seconds under authentication mode.