Touch Analysis: An Empirical Evaluation of Machine Learning Classification Algorithms on Touch Data (2311.14195v1)

Abstract: Our research aims at classifying individuals based on their unique interactions on touchscreen-based smartphones. In this research, we use Touch-Analytics datasets, which include 41 subjects and 30 different behavioral features. Furthermore, we derived new features from the raw data to improve the overall authentication performance. Previous research has already been done on the Touch-Analytics datasets with the state-of-the-art classifiers, including Support Vector Machine (SVM) and k-nearest neighbor (kNN), and achieved equal error rates (EERs) between 0% to 4%. Here, we propose a novel Deep Neural Net (DNN) architecture to classify the individuals correctly. The proposed DNN architecture has three dense layers and uses many-to-many mapping techniques. When we combine the new features with the existing ones, SVM and kNN achieved the classification accuracy of 94.7% and 94.6%, respectively. This research explored seven other classifiers and out of them, the decision tree and our proposed DNN classifiers resulted in the highest accuracy of 100%. The others included: Logistic Regression (LR), Linear Discriminant Analysis (LDA), Gaussian Naive Bayes (NB), Neural Network, and VGGNet with the following accuracy scores of 94.7%, 95.9%, 31.9%, 88.8%, and 96.1%, respectively.

• The paper introduces a novel Deep Neural Network that outperforms traditional classifiers by achieving 100% accuracy on touch data authentication.
• It rigorously compares multiple machine learning models, including SVM, k-NN, and decision trees, using the comprehensive TouchAnalytics dataset of 41 users.
• The study demonstrates the potential of touchalytics for continuous biometric authentication, offering a pathway toward non-intrusive mobile security solutions.

Introduction

The paper explores the field of behavioral biometrics, focusing on unique interactions users have with touchscreen-based smartphones. The research taps into the 'TouchAnalytics' datasets, which consist of behavioral features from 41 subjects. These features are pivotal in differentiating one individual from another based on their interaction with touch screens. The paper builds upon existing research wherein classifiers like Support Vector Machine (SVM) and k-nearest neighbor (kNN) were used and achieved very low equal error rates, indicating high performance. The innovation of this paper lies in proposing a novel Deep Neural Network (DNN) architecture for classifying individuals based on touch data, aiming to enhance authentication performance.

Touchanalytics and Behavioral Biometrics

Touch-based interactions, known as touchalytics, is an emerging player in the field of behavioral biometrics. Unlike physical characteristics such as fingerprints and facial recognition, which are commonly used in authentication systems, touchalytics harnesses unique patterns in touchscreen interactions. This research assesses the viability of touchalytics as a standalone authenticator, offering continuous authentication that can signal security systems about the person currently interacting with the device—a level of ongoing security not afforded by traditional biometric methods.

Machine Learning Algorithms for Touch Data

The paper conducts a comprehensive analysis of various machine learning classifiers applied to touch data. Decision trees and the proposed DNN outperformed the rest, showing a perfect classification accuracy of 100%. Other classifiers evaluated include Logistic Regression, Linear Discriminant Analysis, Gaussian Naive Bayes, Neural Network, and a modified version of VGGNet. The paper also explores using a Genetic Algorithm (GA) for feature selection, aiding in pinpointing which data points are most influential for classification accuracy. This optimization shows promise for yielding even better results when applied to a large number of features.

Datasets and Outcomes

The research utilizes two datasets involving 41 users interacting with five different Android phones. Authentication accuracy was assessed using raw user data with fewer features and extracted data with a more comprehensive set of features. Results indicate that while the SVM and k-NN classifiers perform well, the newly proposed DNN classifier brings accuracy to an even higher level. This proficiency suggests that touchalytics, empowered by advanced machine learning techniques, has strong potential as a method for user authentication.

Conclusion and Implications

In conclusion, this research opens promising avenues for the application of touchalytics in user authentication, presenting it not only as a complementary security measure but also highlighting its potential for continuous authentication. The high accuracy rates achieved by the proposed DNN classifier place it at the forefront of touch data classification. The findings may influence the development of more secure, non-intrusive authentication systems, which could fundamentally transform personal device security. Future studies aim to refine these methodologies by investigating classifier performance on an individual user basis, to enhance the personalization and reliability of touch-based authentication.