Federated Re-Identification With Mapping for Non-Overlapping Domains
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing federated learning techniques struggle to create a domain-generalisable model for re-identification of targets in non-overlapping data sets, particularly in visual data representing different person populations from different locations, due to discrepancies in class spaces and privacy concerns with centralised data sharing.
Innovation Solution
Federated Person Re-Identification (FedReID) uses local data from clients to optimise local feature embedding models, with a central server aggregating updates to a global model while maintaining privacy, allowing for domain-generalisable re-identification without sharing local data, and incorporating a mapping network to align relevant updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federated learning is used to train a centralised model from local data, then privacy is preserved, but the model cannot handle non-overlapping class spaces across different domains
Solution Approach 1:
The system segments the learning process into local model training at each client and centralized aggregation at the server. Each client trains a local model on its private data without sharing the data itself, while the server aggregates these local models to create a domain-generalisable global model. This segmentation allows privacy preservation at the client level while achieving domain adaptability through the aggregated global model.
Solution Approach 2:
The patent introduces an intermediary mapping layer that translates between different local class spaces and a unified global class space. This mapping mechanism acts as a mediator that enables the global model to handle non-overlapping class spaces across different domains while maintaining the privacy benefits of federated learning, as the mapping is learned and exchanged without requiring raw data sharing.
2Measurement precision
If local data is shared and centralised for training, then a better global model is obtained, but privacy concerns and data protection requirements increase
Solution Approach 1:
The system extracts only the essential model parameters and gradients needed for training from the local data, while leaving the actual sensitive data at the client side. This extraction approach allows the global model to benefit from aggregated knowledge across multiple domains without requiring access to the private local data itself, thereby maintaining model accuracy while eliminating privacy risks associated with data centralisation.
Solution Approach 2:
Instead of sharing original data, the system creates and exchanges copies of model parameters, gradients, and learned representations. These copies contain the essential learning information needed to improve the global model while being mathematically transformed versions that do not reveal the underlying private data, thus achieving model improvement without privacy compromise.
3Reliability
If existing federated learning techniques are applied to non-overlapping data sets, then local privacy is maintained, but discrepancies in class spaces prevent effective model aggregation
Solution Approach 1:
The patent transforms the class space representation through learned mapping parameters that adapt local class spaces to a unified global class space. These parameter changes are dynamically learned during the federated training process and allow the system to handle non-overlapping class spaces without compromising privacy, as the transformation is achieved through parameter exchange rather than data sharing.
Solution Approach 2:
The system introduces an additional dimensional layer in the form of mapping networks that bridge different class spaces. This dimensional transformation allows the global model to accommodate diverse local class spaces by projecting them into a unified representation space, effectively managing the complexity of class space alignment while maintaining the privacy benefits of federated learning.
Data Source
AI summary
A method for generating an optimised domain-generalisable model for re-identification of a target in a set of candidate images. The method optimises a local feature embedding model for domain-specific feature representation at each client of a plurality of clients, then receives, at a central server, information on changes to the local feature embedding model at each respective client resulting from the optimising step, and then updates a global feature embedding model based on the changes to the local feature embedding model. The method further receives, at each client from the central server, information representative of the updates to the global feature embedding model, then maps, at each client, on to the respective local feature embedding model at least a portion of the received updates, and subsequently updates, at each client, the respective local feature embedding model based on the mapped updates. The steps are repeated until convergence criteria are met, wherein the global feature embedding model is the optimised domain-generalisable model for re-identification of a target in a set of candidate images.