Biometric Anti-Spoofing with Embedding Similarity and Early Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing biometric systems, particularly fingerprint recognition, struggle to differentiate between genuine and forged biometric information, leading to security vulnerabilities due to the inability to accurately detect spoofing attempts using conventional neural networks that are not adaptable to environmental changes or user-specific variations.
Innovation Solution
An anti-spoofing apparatus utilizing a dual-stage shallow deep neural network (DNN) system that performs early decision classification through (1-1) and (1-2) classifiers to determine forgery scores from intermediate layers, combined with similarity calculations based on embedding vectors and threshold ranges, and updates models dynamically to adapt to environmental changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional neural networks are used for biometric recognition, then the system is simple to implement, but the system cannot accurately detect spoofing attempts and is not adaptable to environmental changes
Solution Approach 1:
The patent divides the neural network into multiple stages: a first shallow DNN for initial forgery detection and a second deep DNN for final recognition. This segmentation allows the system to perform early decision classification, improving spoof detection accuracy while managing computational complexity by handling easy cases in the first stage.
Solution Approach 2:
The system performs preliminary anti-spoofing detection before final biometric recognition using the first shallow DNN. This preliminary action filters out obvious spoofing attempts early in the process, improving overall reliability while reducing the computational burden on the deeper network components.
2Measurement precision
If end-to-end deep neural networks are used, then detection accuracy may be improved, but processing speed decreases and computational resources are consumed
Solution Approach 1:
By segmenting the detection process into two stages with different network depths, the system achieves a balance between accuracy and speed. The first shallow DNN quickly processes obvious cases, while the second deep DNN handles more complex cases that require deeper analysis, optimizing overall processing efficiency.
Solution Approach 2:
The system applies partial deep processing only when necessary - using the first shallow DNN for initial filtering and resorting to the second deep DNN only for cases that require more sophisticated analysis. This partial application of deep processing maintains high accuracy for critical cases while preserving processing speed for the majority of cases.
3Adaptability or versatility
If fixed neural network models are used, then the system is stable, but the system cannot adapt to terminal changes or fingerprint variations
Solution Approach 1:
The patent implements dynamic model updating mechanisms where the neural network parameters are adjusted based on environmental conditions, terminal characteristics, and fingerprint variations. This allows the system to adapt to changing conditions while maintaining a relatively simple update mechanism that builds upon the existing dual-stage architecture.
Solution Approach 2:
The system incorporates feedback loops that monitor detection performance and environmental changes, using this information to dynamically adjust the neural network models. This feedback mechanism enables adaptation to terminal and fingerprint changes while maintaining system stability through controlled, data-driven updates.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
An anti-spoofing method includes detecting first information related to whether the biometric information is forged, based on a first output vector of a first neural network configured to detect whether the biometric information is forged from the input data, extracting an input embedding vector including a feature of biometric information of a user from input data including the biometric information, calculating a similarity value of the input embedding vector based on a fake embedding vector and either one or both of a real embedding vector and an enrollment embedding vector that are provided in advance, calculating a total forgery score based on the similarity value and a second output vector of the first neural network according to whether the first information is detected, and detecting second information related to whether the biometric information is forged, based on the total forgery score.