Abstract
This paper presents a robust, implicit user identification system for smartphones based on gait recognition. By fusing data from four built-in sensors—accelerometer, linear accelerometer, gyroscope, and magnetometer—and employing a novel non-linear feature selection method called Extended Sammon Mapping Projection (ESMP), the system achieves identification accuracies up to 99%. This approach provides a continuous, unobtrusive layer of security that remains effective across various phone placements and walking conditions.
Introduction
Traditional smartphone security relies on explicit identification methods like PINs or biometrics (fingerprints, facial recognition) performed at the point of entry. However, these methods do not protect a device if it is snatched while unlocked. Implicit identification offers a solution by continuously verifying the user's identity in the background. Gait-based identification is a promising biometric, but it often suffers from "within-class variance" caused by different phone positions (e.g., front vs. back pockets) and walking speeds. This research addresses these challenges through multi-sensor fusion and advanced feature selection.
Methodology
The proposed identification pipeline consists of several key stages:
- Multi-Sensor Fusion: The system integrates data from the Accelerometer (AC), Linear Accelerometer (LAC), Gyroscope (GY), and Magnetometer (MM) to capture a comprehensive signature of the user's movement.
- Preprocessing and Feature Extraction: Raw signals are filtered and segmented into windows (5 to 105 samples). From each window, 180 features are extracted, encompassing time-domain, frequency-domain, and specific gait characteristics.
- Extended Sammon Mapping Projection (ESMP): As the core innovation, ESMP is used for non-linear unsupervised feature selection. It reduces the high-dimensional feature set into a lower-dimensional biometric representation while preserving the structural distances between data points. This helps the system distinguish unique gait patterns from the noise of phone placement.
- Classification: The refined features are processed by various machine learning models, including Kernel Extreme Learning Machine (KELM), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Bagging.
Experimental Setup
The study utilized data from 16 individuals (8 male, 8 female). Participants walked for 10 minutes with a smartphone (LG Nexus 4) placed freely in one of four pants pockets. To ensure real-world applicability, the dataset included variations in walking speeds and clothing styles (e.g., different types of jeans).
Results and Discussion
The experimental results demonstrate the high efficacy of the proposed system. The choice of classifier and window size significantly impacts performance.
Accuracy and Model Performance
The system achieved near-perfect accuracy as the complexity of the models increased. For instance, the KELM model's accuracy stabilizes at approximately 99% as the number of hidden neurons increases, showing strong generalization between training and test sets.
The Impact of Window Size
A critical finding is the trade-off between identification accuracy and latency. Larger window sizes provide more data for classification, leading to higher accuracy but increasing the time required for a verdict.
As shown in the confusion matrices for the Bagging classifier, increasing the window size from 5 to 105 samples dramatically improves the True Positive Rate. At 105 samples, the system achieves nearly 100% accuracy across all 16 users, effectively eliminating false negatives.
Conclusion
The combination of multi-sensor fusion and ESMP feature selection provides a highly accurate and position-independent method for implicit smartphone security. While the system shows great promise, future work should address temporal consistency (how gait changes over months or years), hardware heterogeneity across different smartphone manufacturers, and further computational optimization for real-time mobile deployment.
Summary updated