When modeling the human hand, features such as radius or origin are difficult to imitate in our method. Most applications run on mobile devices like desktop applications, as mobile computing power continues to increase. Although unlocking methods such as password, PIN, and geometric pattern lock have been widely adopted, it is vulnerable to shoulder surfing attacks because they are subjective.
We collect 5 values from smartphone sensors such as touch screen and gravity sensor and extract 25 features.
Raw Data Process
Android API (Application Programming Interface) should be used to get touch data from touch screen and gravity sensor. Android API catches touch data 17~18ms generally and gravity sensor value with certain interval.
Features 1~3: Radius and Origin
Definition
Extraction Method
Feature Analysis
Feature 4: Length of Drag
Definition
Extraction Method
The value of the gravity sensor is able to control how the mobile phone is positioned. The location and status of the mobile phone is therefore an important factor in separating the people. These figures show that the shape of the pulling movement is different for each people.
One of the important techniques of the SVM is to solve nonlinear classification problems using the kernel. In the case where K is 1, it depends only on the data of the nearest, it is sensitive to noise. Instead of finding the right answer, there is interest in discovering the structure of the data set.
Because VQ is not a class designation and is one of the unsupervised learning models. It is a gesture that uses the forearms and wrist, which is different from the gesture above. In general, the value of the gravity sensor is similar until people hold the mobile phone in the posture.
Feature 5~8: Start and End Point of Drag
Definition
If people extend their finger on a mobile phone, the start and end points of the swipe are approximately the same because the shape of the finger does not change. This is the speed that is created when you create a drag until you grab the cell phone. a) Figure 8 shows the graph of the -coordinate drag speed. However, in the H2 cases, the performance is not as good compared to the data distribution.
It makes it possible by implementing a slack variable and solves to some extent in the case of linear separable problem, but it is limited because it uses as a classification bounds for the linear hyperplane. If the assumption dimension of the kernel is very high, computational sum of the high-dimensional vectors is not practical. The more the graph is drawn near the top of the left, the more the classification performance is superior.
Third, we present the passage of FNR (False Negative Rate) and TPR through SVM threshold adjustment. SBS is an optimization technique for automatic feature selection when n number of features are given. The performance of the classifier can be reduced by redundant features if the feature space is large.
The weight of the distance corresponding to the range of each feature is not taken into account, but only the Euclidean distance is used. The experiments are conducted with four positions because we guess the value of the feature change when using a mobile phone in different poses. When you use a mobile phone with a leaning or lying posture, the value of the function changes because the position of the arm and wrist position is different.
The patterns of the drags differ from person to person because the shape of the hand does not change.
Feature Analysis
Features 9~11: Gravity
Feature 12~17: Touch Area
These are determined by the length of the longest and shortest part of the contact area.
Feature 18~23: Velocity
Definition
Extraction Method
Feature Analysis
Feature 24~25: Perpendicular
Definition
Extraction Method
Since learning is the most important element that humans have, one makes a decision through learning. In a field of artificial intelligence, machine learning is an algorithm for a computer to be able to learn. The machine learning algorithm, when the ability to enter training data into a computer, performs to predict what the class type will be based on the algorithm of any particular criterion of building a discrimination criterion.
Support Vector machine
Support vector machine is a supervised learning model used for regression and classification in machine learning. SVM is that the method of classifying the data into two categories is to find the optimal hyperplane that separates the two groups as far as possible for the given data. Since there are many hyperplanes to classify the data generally belonging to two categories, SVM can find which hyperplanes are optimal.
In order for the computed hyperplane to have a boundary to separate the most distant two classes of the given data, it should make the hyperplane that maximizes the margin. The SVM hyperplane defined as an edge is the distance between the data points on these lines, bisects the middle of the edge, which is the largest size, and the point consisting of lines is called the support vector. In this case, any data belonging to a class above (1) or below (2) must satisfy the following conditions to be located.
To calculate the hyperplane with the maximum edge, even if another class is present in the function between them, a slack variable is added to the equation. By mapping the dimension of the input data to the dimension ϕ(x) used for dimensionality scaling, nonlinear classification problems can be solved using a simple linear classifier. SVM implements the trick kernel method to solve the problem of these operations.
K-Nearest Neighbor
In k-NN, it is necessary to take time to calculate the distance of all training data when receiving new data and select k adjacent data. Therefore, the k-NN algorithm needs the large storage capacity and the computation problem that takes a long time to compute. If K is large, it becomes possible to reference the other class, therefore, the result is determined according to the ratio data of each occupied class.
Vector Quantization
Ward's method uses the distance between two clusters, k and j, and indicates how much the sum of squares will increase if we merge them. Luca's method uses dynamic time warping (DTW) to analyze a time series of touchscreen data, such as all combinations of x coordinates, y coordinates, pressure, size and time. In terms of safety, the false positive rate is very important because a false positive case is considered a serious problem in terms of safety. Their work shows that the false positive rate is too high, about 21%.
They measured parameters of SVDE such as γ, parameter of RBF and ν, a parameter of SVDE by grid search with 10 cross-validation on each training group. Saevanee proposed password user authentication by using additional information such as interval between sequential touch, duration and pressure generated when the user touch on the touch screen [12]. Saevanee's method works in the notebook touchpad and captures signals such as measuring force and value of keystroke dynamics when the user enters a 10-digit number on the touchpad.
In Meng's method, multi-touch is applied to process a number of fingers on a touch screen at the same time as Shahzad's proposed method. While the user is holding the mobile performing the gesture, the user can touch the screen with the thumb. They have the main objective of capturing a biometric amount of muscle memory and physical characteristics of the specific user using gesture.
Experimental Setting
Accuracy Evaluation
It is possible to monitor the increase in TPR as the CV number increases. Sequencing is a method of increasing the TPR and decreasing the FPR to prevent the SVM output from accidentally being a false positive. In the experiment it is performed in the Samsung Galaxy Note2 on Android. To create the models, two people created 10 to 15 minutes of training data.
When it comes to looking in the ROC curve, collecting false positive data improves performance. Therefore, it is necessary to define the number of first-time unlock attempts (i.e., false positives) that occurred when a fraudster tried to unlock rather than being defined in relation between FPR and TPR. If the threshold is adjusted to 0.3 to reduce the occurrence of FP, the unlock occurred with a probability of 2.78% when cheaters tried at least 20 times.
Because the experiment was conducted in the same size phone on the Galaxy Note2, we did not consider the screen size in this experiment. Methods such as taking the size into account are necessary for it to be implemented on other devices. In this thesis, we introduced classification, performance evaluation and the DragID authentication system to analyze the pattern of drag to model the people.
In the real-world experiments, we showed good performance on shoulder surfing attacks using the 10-15 minute training data. Liu, “Securely unlocking mobile touchscreen devices by simple gestures – you can see it, but you can't do it,” in Proc, ACM(Mobisys), 2013.
Real-world Evaluation
Comparison with Classification
Comparison with Existing Schemes
Limit of Posture
Size of Phones
To prove this, pulls can represent people from the first attempt. In another experiment, it was shown that the posture when picking up the phone is almost the same over time. We expect DragID to have more performance by collecting data and updating the model every time users unlock the app.
Global mobile statistics 2014 Part A: Mobile subscribers; mobile phone market share; mobile operators”, http://mobithinking.com/mobile-marketing-tools/latest-mobile-stats/a#subscribers. Chaos Computer Club breaks Apple TouchID”, http://www.ccc.de/en/updates/2013/ccc- breaks-apple-touchid. Saevanee, "User authentication through combination of behavioral biometrics over the touch panel acting like touch screen of mobile device", ICCEE, 2008.
Dynamic Update Learning