Abstract
MobileTouchDB is a large-scale, "in-the-wild" database of touch-based handwriting collected from 217 users across 94 smartphone models. This paper introduces the dataset and benchmarks a behavioral biometric system using Dynamic Time Warping (DTW). Results indicate that analyzing neuromuscular patterns in touch strokes can significantly enhance mobile security, achieving an Equal Error Rate (EER) of 5.9% for 9-character passwords and providing a robust second factor against traditional authentication attacks.
Introduction
Traditional mobile authentication—including PINs, patterns, and passwords—is increasingly fragile. These methods are susceptible to "smudge attacks," where oily residue on the screen reveals patterns, and "shoulder surfing," where observers watch a user enter their credentials. To address these vulnerabilities, researchers are shifting toward behavioral biometrics. Instead of merely verifying what a user knows, these systems verify how a user interacts with their device, leveraging unique neuromuscular patterns.
However, progress in touch-based biometrics has been hindered by the lack of realistic, large-scale datasets. Most existing data is collected in controlled laboratory settings. This paper introduces MobileTouchDB, a dataset designed to provide a realistic benchmark for mobile touch biometrics in unsupervised, "in-the-wild" conditions.
Methods and Data Collection
The researchers developed an Android application distributed via the Google Play Store, enabling users to contribute data using their own devices in their natural environments.
Dataset Characteristics
MobileTouchDB is significantly more diverse than previous benchmarks:
- Scale: Over 64,000 on-line character samples.
- Diversity: 217 users across 94 different smartphone models.
- Content: A comprehensive set including numbers (0-9), upper and lower-case letters (54), symbols (8), and specific 4-digit passwords.
- Protocol: To capture "inter-session variability"—how handwriting changes over time—data was collected over six sessions per user, with at least a two-day gap between sessions.
Biometric Evaluation
To evaluate the dataset, the authors implemented a baseline biometric system using Dynamic Time Warping (DTW). For each character, the system extracts six time-based functions: X and Y coordinates, velocity (first-order derivatives), and acceleration (second-order derivatives). DTW aligns a query sample against a stored template to minimize the distance between sequences, accounting for variations in speed and stroke length.
Key Insights and Results
The study revealed several critical factors influencing the effectiveness of touch biometrics:
- Complexity and Discriminability: Symbols (e.g.,
%,*) proved more discriminative than numbers because their complex stroke patterns provide a larger "biometric surface area." Similarly, lower-case letters yielded a 1.0% improvement in Equal Error Rate (EER) over upper-case letters, likely due to more individualized curves and flourishes. - Character-Specific Performance: Not all characters are equally secure. For instance, the number "8" was the most discriminative (22.6% EER) due to its complex loops, while "6" was the least.
- The Power of Length: Security scales with password length. The EER dropped from 19.5% for a single character to 5.9% for a 9-character password.
Benchmarking
The MobileTouchDB approach significantly outperformed existing standards. On a 9-digit password task, it achieved a 5.9% EER, compared to the 9.0% EER reported for the e-BioDigit benchmark. Crucially, even for common passwords like "123456," the biometric layer maintained a 17.0% EER against an impostor who already knows the password—a scenario where traditional systems would have a 0% security margin.
Discussion and Future Work
While the DTW baseline is effective, the authors suggest that Deep Learning architectures, such as Recurrent Neural Networks (RNNs) or LSTMs, could further refine accuracy by capturing more subtle temporal nuances. Future iterations of MobileTouchDB will incorporate auxiliary sensor data—such as accelerometer and gyroscope readings—to capture how users hold and move their devices during authentication. These additions aim to create an even more robust and multi-dimensional biometric signature.
Summary updated