TL;DR
This paper introduces a fuzzy expert system for liveliness detection in biometric authentication, aiming to combat spoofing attacks. By combining eye/mouth movement and image texture analysis, the system demonstrates robustness, particularly under challenging illumination conditions, achieving a False Rejection Rate (FRR) of 0.28 in such scenarios.
What Problem Does This Paper Address?
Face recognition systems are vulnerable to spoofing attacks where an attacker uses artifacts like photos, videos, or masks to impersonate an enrolled user. Traditional face recognition systems often overlook the "liveliness" of the presented face, focusing solely on matching facial features. While various liveliness detection methods exist, including challenge-response, texture analysis, and multi-modal approaches, they often struggle with variable illumination or require user interaction. This paper addresses the need for a robust, user-friendly, and automated liveliness detection system that can effectively handle real-world conditions, especially poor lighting.
Methods
The authors propose a fuzzy expert system based on the Mamdani inference style. The system takes three inputs:
- Eye Movement: Detected using Haar cascade classifiers and analyzed for movement using Scale Invariant Feature Transformation (SIFT).
- Mouth Movement: Similar to eye movement, detected and analyzed using Haar cascades and SIFT.
- Image Quality: Assessed using Local Binary Patterns (LBP) to measure surface texture and homogeneity, which differs between real faces and glossy artifacts.
These inputs are fuzzified into linguistic variables (e.g., "poor," "average," "good") using trapezoidal membership functions. A rule base, defined using AND/OR operators, then infers the liveliness. Finally, the fuzzy output is defuzzified using the Center of Gravity (COG) method to produce a crisp decision: "Person is Live" or "Person is Not Live."
Datasets
The paper mentions using the following datasets:
- Real Faces: A combination of faces from the Robotics and Artificial Intelligence Laboratory at the Indian Institute of Information Technology, Allahabad, and the University of Essex, UK (face94) dataset.
- Fake Faces: Created using artifacts such as photos on a laptop, 2D plain paper, and high-definition video from an HD mobile phone, utilizing the real face datasets for these attacks.
Key Insights and Ideas
- Fuzzy Logic for Robustness: Fuzzy logic is particularly well-suited for handling the inherent uncertainty and variability in real-world biometric data, especially under non-ideal conditions like poor illumination.
- Multi-faceted Liveliness Cues: Combining eye/mouth movement with image texture provides a more comprehensive approach to liveliness detection than relying on a single cue. Real faces have distinct texture properties (roughness) compared to artificial surfaces.
- Illumination Invariance: While SIFT is partially invariant to illumination, the fuzzy system's ability to map inputs to linguistic terms like "poor" or "average" allows it to adapt to degraded image quality caused by poor lighting.
- FRR as a Key Metric: The paper emphasizes False Rejection Rate (FRR) as a critical performance indicator, especially in challenging scenarios, and aims to minimize it.
Key Results
The proposed fuzzy expert system demonstrates strong performance:
- Ideal Conditions: Achieves 100% classification accuracy against all tested attack types (laptop photo, 2D paper photo, HD video).
- Testing Parameter-I (Ideal Case): Not explicitly detailed with numbers, but implied to be high accuracy.
- Testing Parameter-II (Bad Illumination): Achieves an FRR of 0.28.
- Testing Parameter-III (No/Very Less Movement): Achieves an FRR of 0.4.
- Attack Scenarios: For all three attack types (laptop, 2D paper, HD video), the system shows very good accuracy with zero False Acceptance Rate (FAR).
Limitations and Future Work
- Limitations:
- The paper acknowledges that in bad illumination, the system's ability to detect eye and mouth movement is hindered, impacting accuracy.
- The FRR of 0.4 in scenarios with minimal movement suggests room for improvement.
- Future Work:
- The authors suggest that tuning parameters like input fuzzification, linguistic variable ranges, membership functions, and rule database settings can further improve performance.
- They propose using Equal Error Rate (EER) for selecting the decision threshold to discriminate between real and imposter.
- The flexibility of the framework allows for integration with existing literature to achieve significant classification improvements.
Summary updated