Federated Learning System for Biometric Authentication Fairness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current biometric authentication systems, particularly face authentication, face challenges in maintaining accuracy across different attributes such as gender and skin color, leading to disparities in authentication accuracy and false recognition rates, which existing methods have not adequately addressed.
Innovation Solution
A learning system that employs associative learning by updating models on client devices and a training server, where individual models are generated and optimized based on classification results, reducing disparities in authentication accuracy by considering attributes like gender and skin color through a common model and individual model framework.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a common face authentication model is trained using deep learning, then authentication accuracy is greatly improved, but a large amount of face images is required for learning which becomes difficult to collect due to privacy protection systems
Solution Approach 1:
The learning system is segmented into multiple client devices that each perform local learning on their own data, rather than requiring centralized collection of all data. Each client device trains a local model on its own face images, and only model parameters are shared and aggregated at the server level, eliminating the need to collect large amounts of personal face images centrally while still achieving high authentication accuracy through federated learning
Solution Approach 2:
The server acts as an intermediary that coordinates the federated learning process by distributing the common model to client devices, collecting updated model parameters from each client, and aggregating them to update the common model. This intermediary structure enables collaborative learning without direct data sharing between clients or central collection, resolving the privacy-accuracy tradeoff
2Productivity
If a face authentication model is trained with general data, then the model can be updated efficiently, but a difference in authentication accuracy occurs between individuals having different attributes such as gender or skin color
Solution Approach 1:
The system applies local quality by allowing each client device to perform local learning on its own data with its own characteristics (gender, skin color, etc.), generating local model updates that are specific to each client's data distribution. This ensures that minority attribute groups contribute their specific characteristics to the global model, improving fairness while maintaining overall efficiency through parallel local learning processes
Solution Approach 2:
The system changes parameters by adjusting the aggregation weights of model updates from different clients based on their data characteristics and contribution quality. By dynamically modifying aggregation parameters rather than using uniform averaging, the system can compensate for imbalances in attribute representation and improve fairness across different demographic groups while maintaining efficient model updates
Data Source
AI summary
A training server classifies a common model and an individual model transmitted from a plurality of client devices on the basis of the individual model transmitted from a training data management server, and updates the common model and the individual model in accordance with a classification result.