TL;DR
This paper introduces PUPGUARD, a novel two-factor authentication system that combats "puppet attacks" by analyzing both fingerprint image features and the dynamic timing patterns of user interactions. By requiring two distinct finger presses with a specific time interval, PUPGUARD significantly enhances security against coerced authentication attempts, achieving an impressive 97.87% accuracy.
What Problem Does This Paper Address?
Traditional fingerprint authentication systems, while convenient and generally secure, are vulnerable to "presentation attacks" where fake fingerprints are used. A more insidious threat is the "puppet attack," where an attacker forces a legitimate user to authenticate against their will, often through coercion or intimidation. In these scenarios, the fingerprint itself is genuine, making it difficult for standard systems to detect the illegitimate intent. Existing defenses against puppet attacks are limited, often requiring manual guidance or being impractical for stationary devices. This paper addresses the critical need for an automated and robust defense against puppet attacks that can distinguish between a genuine user acting freely and a coerced user.
Methods
PUPGUARD employs a two-factor authentication approach that goes beyond static fingerprint data. The core idea is to capture and analyze two distinct behavioral patterns:
- Fingerprint Image Features: Standard image processing techniques are used to extract features from the fingerprint itself.
- Timing Characteristics: Crucially, the system measures the time interval between two successive finger presses. This is based on the hypothesis that a coerced user, or an attacker manipulating a user, will exhibit different timing patterns (e.g., hesitation, unusual speed, or difficulty in aligning fingers) compared to a user acting naturally.
The process involves:
- Data Acquisition: A custom system was built to collect fingerprint images and precise timing data from users performing a specific two-press action.
- Preprocessing: Both image and timing data are preprocessed. Image preprocessing includes techniques like Otsu's thresholding and resizing/cropping/normalization. Timing data is standardized.
- Feature Extraction: For image features, the paper explores Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG), and importantly, deep learning features from ResNet architectures (ResNet34, ResNet50, ResNet101, ResNet152).
- Feature Selection: Principal Component Analysis (PCA) is used to reduce the dimensionality of the image features.
- Feature Fusion: The extracted image features and the timing characteristics are fused into a single feature vector.
- Classification: A one-class classifier (specifically OC-SVM, Isolation Forest, and LOF are explored) is used to distinguish between normal and anomalous behavioral patterns.
The paper also investigates decision-level fusion as an alternative to feature fusion.
Datasets
The paper introduces a newly created dataset for training and evaluation. This dataset comprises:
- 496 fingerprint pairs (992 individual fingerprints) and their associated time intervals.
- Collected from 31 individuals aged between 20 and 85.
- The data was collected under normal user conditions, meaning it does not contain direct examples of puppet attacks but rather the baseline behavior against which attacks are detected.
- A separate testset of 94 fingerprint pairs (188 fingerprint images) was created, including simulated abnormal behaviors (forceful pressing, unusual timing) to represent attack scenarios.
Key Insights and Ideas
- Behavioral Biometrics are Key: The paper strongly advocates for the use of dynamic behavioral patterns (timing) in addition to static biometric traits (fingerprint images) for enhanced security.
- Timing is a Powerful Indicator of Coercion: The core insight is that the time interval between two distinct finger presses is a sensitive indicator of a user's state, particularly when under duress. A coerced user's natural resistance or the attacker's manipulation will likely alter this timing.
- Two-Factor Authentication for Behavior: By requiring two distinct finger presses, PUPGUARD effectively creates a "behavioral two-factor authentication," making it significantly harder for attackers to mimic.
- Deep Learning for Image Features: The use of ResNet architectures proved crucial for extracting robust image features, outperforming traditional methods like LBP and HOG.
- Feature Fusion is Superior: Combining image and timing features (especially through "feature cross") significantly outperforms using either feature type alone, highlighting the complementary nature of these data sources.
Key Results
- High Accuracy and Low FPR: The proposed PUPGUARD system achieved a remarkable 97.87% accuracy and a very low 1.89% False Positive Rate (FPR) when using ResNet50-based features with OC-SVM and feature cross-fusion.
- ResNet Dominance: ResNet-based features consistently outperformed LBP and HOG features across various classifiers and fusion methods.
- Feature Fusion Superiority: Using both image and timing features through fusion methods (especially feature cross) yielded significantly better results than using image features alone (max accuracy ~70%) or timing features alone (max accuracy ~95%, but with higher FPR and risks).
- Same Finger Pressing is Vulnerable: Experiments showed that allowing the same finger to be pressed twice resulted in significantly lower accuracy and higher FPR, underscoring the importance of using distinct fingers.
- Dataset Size Matters: Increasing the training dataset size led to a stabilization and improvement in detection accuracy.
Limitations and Future Work
- User Adoption and Usability: The requirement for users to perform two distinct finger presses might introduce complexity, potentially affecting user adoption and overall usability.
- Implementation Constraints: Adapting existing hardware and software to incorporate dynamic behavioral pattern analysis can pose technical challenges.
- Dataset Limitations: While a new dataset was created, it primarily consists of normal user data. Future work could involve collecting more diverse and realistic puppet attack scenarios.
- Further Exploration of Fusion: The paper mentions exploring decision-level fusion and other fusion techniques, suggesting avenues for future research.
- Real-World Deployment: Testing PUPGUARD in diverse real-world environments and with a wider range of users and devices would be beneficial.
Summary updated