Cascade Liveness Detection for Known and Unknown Spoof Attacks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Facial recognition systems are vulnerable to presentation attacks and existing monolithic threat detection models are inefficient in detecting both known and unknown threats, requiring extensive training and retraining to maintain detection efficacy.
Innovation Solution
A cascaded system comprising filters, special models, and a general model ensemble for liveness detection, performing multiple levels of security checks to effectively identify and classify threats, including unknown threats, with minimal computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a monolithic threat detection model is used, then the system structure is simple, but the detection accuracy for different types of threats deteriorates
Solution Approach 1:
The monolithic threat detection model is segmented into multiple specialized models, each trained to detect specific types of threats (e.g., photos, videos, masks, synthetic fingerprints). This segmentation allows each model to specialize in detecting particular attack vectors, thereby improving overall detection accuracy while maintaining manageable system complexity through modular architecture.
Solution Approach 2:
The patent implements a universal threat detection system that combines multiple specialized models under a single framework. The system can detect various types of threats including photos, videos, masks, and synthetic fingerprints using different specialized models, making the system multi-functional and adaptable to diverse attack vectors without requiring separate dedicated systems for each threat type.
2Measurement precision
If multiple specialized models are built for each threat type, then the detection accuracy for known threats improves, but the device complexity and training requirements deteriorate
Solution Approach 1:
The threat detection system is divided into multiple specialized models, each trained for specific threat types. This segmentation enables high detection accuracy for known threats while managing complexity through modular design, where each model can be independently trained and maintained.
Solution Approach 2:
The system dynamically adjusts the number and type of specialized models activated based on the detected threat characteristics. Rather than always running all possible models, the system changes parameters to select only the relevant specialized models for each detection scenario, reducing computational overhead and training requirements while maintaining high accuracy.
3Measurement precision
If a cascade of specialized models is implemented, then the detection accuracy for known threats improves, but the ability to detect unknown threats deteriorates
Solution Approach 1:
The system segments threat detection into specialized models for known threats and a separate general model for unknown threats. This segmentation allows specialized models to achieve high accuracy for known attack vectors while the general model maintains versatility in detecting previously unknown threats, resolving the trade-off between specialization and adaptability.
Solution Approach 2:
The general model acts as an intermediary that receives inputs from specialized models and provides a final detection verdict. This intermediary component ensures that unknown threats are not missed by specialized models, as the general model can detect patterns that don't match any specific known threat type, thereby maintaining both accuracy for known threats and versatility for unknown threats.
4Measurement precision
If specialized models are retrained frequently to maintain detection efficacy, then the detection accuracy improves, but the time and computational resources required deteriorate
Solution Approach 1:
The system changes parameters to select and activate only the necessary specialized models for current detection scenarios rather than retraining all models. This allows the system to maintain high detection efficacy by updating only relevant models while avoiding the time-consuming process of retraining the entire model suite, thus reducing computational overhead and time loss.
Data Source
AI summary
Systems and methods for performing liveness detection. A method includes transmitting an image or a video by a threat detector to one or more filters for a first level check for detecting a threat based on analysis of certain characteristics, transmitting the image or the video to one or more special models for a second level check, wherein the one or more special models are configured for detecting multiple types of threats on the image or the video, transmitting the image or the video to a general model ensemble for a third level check to classify the image or the video according to individual features into original and fake, and in response to detecting the threat, registering the threat by the threat detector.