Introduction
Mobile biometrics are evolving beyond static identifiers like fingerprints to behavioral patterns, such as how a user interacts with a touchscreen. However, analyzing these gestures is challenging due to the "multi-view" nature of the data—where information is captured across different feature sets (e.g., velocity vs. pressure) or sensors. Traditional methods often struggle with high intra-class variability and the "curse of dimensionality," where features outnumber available training samples. This research introduces a novel approach to fuse these disparate data views effectively.
The DFS Framework
To address these challenges, researchers have developed Discriminative Factorized Subspaces (DFS). This method maps high-dimensional multi-view data into a lower-dimensional "factorized" representation. Unlike previous approaches that treat views as either entirely independent or redundant, DFS identifies two distinct types of information:
- Shared Subspace: Captures commonalities across all data views.
- Private Subspaces: Capture unique characteristics specific to individual views.
The technical core of DFS involves a joint optimization problem using a constrained Rayleigh Quotient. By maximizing between-class separation and minimizing within-class scatter, the model ensures that different users are easily distinguishable while a single user's swipes remain consistent. An orthogonality constraint is applied to prevent overlap between shared and private subspaces, forcing a true factorization of information. The solution is found efficiently using generalized eigenvalue decomposition, making it significantly faster than iterative deep learning approaches.
Experimental Evaluation
The DFS model was evaluated against four public touchscreen biometric datasets: Serwadda, Frank, Antal, and UMDAA. These datasets encompass a wide range of users and interaction types, from landscape/portrait modes to unconstrained phone usage over several months. Features extracted included 28 swipe-based metrics (velocity, pressure, curvature) and 5 signature-based metrics (stroke rhythm).
Key Findings
DFS consistently outperformed state-of-the-art methods such as DCA, PSLF, and SMBR across all test scenarios.
- High Accuracy: In identification tasks, DFS achieved accuracies between 97% and 99%. In verification tasks, it significantly reduced Equal Error Rates (EER); for instance, on the Serwadda dataset, DFS reached an EER of ~0.74%, compared to >10% for competing methods.
- Efficiency: DFS is computationally efficient, processing thousands of instances in less than a second. This is vital for mobile biometrics, where real-time performance is required.
- Discriminative Power: The factorization approach prevents unique, useful features in one view from being obscured by noise in another, leading to superior class separation even with limited training data (100 or fewer samples).
Conclusion
DFS represents a significant advancement in feature-level fusion for biometrics. Its ability to perform well with limited training data makes it highly practical for real-world mobile security applications. While a kernelized version (KDFS) was explored to handle non-linearities, the linear DFS proved highly effective for current touchscreen feature sets. Future research will explore extending this framework to other modalities like facial recognition or gait analysis and investigating cross-modal identification tasks.
Summary updated