Biometric Embedding Aggregation for Privacy-Preserving Model Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Biometric identification systems face challenges in maintaining user privacy and security while updating neural networks without requiring re-enrollment, as storing input data as a 'gallery' poses privacy risks and is inefficient.
Innovation Solution
Utilizing transformer networks to convert embeddings between different neural network spaces without storing raw input data, enabling updates and migrations without re-enrollment, and aggregating embeddings from multiple models to enhance information and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If raw input data is stored as a 'gallery' for biometric identification, then identification accuracy is maintained, but user privacy is compromised and storage requirements increase
Solution Approach 1:
The patent extracts only the essential embedding representations from the raw input data while discarding the raw data itself. The embedding data is transformed and stored instead of the original biometric inputs, thereby maintaining identification capability while eliminating privacy risks associated with storing raw data.
Solution Approach 2:
The patent creates a transformed copy of the embedding data in a new embedding space through the transformer network. This copy retains the identification information needed for accurate recognition while being fundamentally different from the original data representation, enabling privacy-preserving storage.
2Measurement precision
If neural networks are updated with new embedding models, then identification accuracy improves, but re-enrollment of all users is required which increases time loss
Solution Approach 1:
The patent introduces a transformer network as an intermediary that bridges old and new embedding spaces. This mediator transforms embeddings from the old model to the new model space, allowing seamless updates without requiring users to re-enroll, thus eliminating time loss while maintaining accuracy improvement.
Solution Approach 2:
The patent changes the embedding space parameters through the transformer network to adapt old embeddings to the new model. This parameter transformation allows the system to leverage existing user data with updated models, avoiding the need for complete re-enrollment and reducing time loss.
3Measurement precision
If multiple embedding models are used to enhance information, then identification accuracy improves, but device complexity increases
Solution Approach 1:
The patent merges multiple embedding models and their transformations into a unified system where the transformer network coordinates different embedding spaces. By combining multiple models through the transformer intermediary, the system achieves enhanced accuracy while managing complexity through a structured integration approach.
Data Source
AI summary
An image of at least a portion of a user during enrollment to a biometric identification system is acquired and processed with a plurality of models to determine embeddings that are representative of features in that image in respective embedding spaces. Transformers are trained to accept as input an embedding from a respective model and produce as output a transformed embedding consistent with a common embedding space. These transformed embeddings are aggregated to produce transformed aggregated embedding data and stored for later comparison to identify the user. Subsequently, one or more embedding models may be used to process a query image to determine a query embedding. The query embedding may be transformed into the common embedding space for comparison.