Personalized Federated Learning for Data-Heterogeneous User Authentication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing federated learning methods for user authentication face challenges with data heterogeneity and loss of local data, compromising security and accuracy, particularly in non-face-to-face financial transactions.
Innovation Solution
A personalized federated learning method that separates the timing for updating global and personalized parameters, using local data to train a first learning model and a second learning model independently, enhancing user authentication accuracy and privacy by applying global parameters to the first model and personalized parameters to the second model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federated learning is used to train learning models locally at user terminals, then user privacy is protected and security is improved, but data heterogeneity arises and local data is lost during global parameter derivation
Solution Approach 1:
The learning model is divided into two distinct models: a first learning model for global parameter training and a second learning model for personalized parameter training. This segmentation allows global knowledge to be shared while preserving local data characteristics and preventing data loss.
Solution Approach 2:
The system implements personalized parameters specific to each user terminal in the second learning model, allowing each local system to maintain its unique data characteristics while benefiting from global learning through the first learning model.
2Quantity of substance
If global parameters are derived from local parameters in federated learning, then collective knowledge is accumulated, but local data distribution differences cause performance degradation
Solution Approach 1:
The authentication system is segmented into a first learning model for global pattern recognition and a second learning model for personalized authentication, allowing each to optimize for its specific function without interference from data distribution differences.
Solution Approach 2:
The system uses only local data characteristics necessary for personalized authentication in the second learning model, rather than attempting to incorporate all local data variations into global parameters, thus avoiding performance degradation from distribution differences.
3Device complexity
If a single learning model is used for both global learning and personalized authentication, then system complexity is reduced, but authentication accuracy deteriorates due to data heterogeneity
Solution Approach 1:
The authentication system is divided into two specialized learning models: the first learning model handles global parameter learning from aggregated local parameters, while the second learning model handles personalized authentication using local data characteristics, with each model optimized for its specific purpose.
Solution Approach 2:
The system adds a new dimension to the learning architecture by introducing personalized parameters in the second learning model that operate independently from the global parameters in the first learning model, allowing simultaneous optimization of both global and local authentication performance.
Data Source
AI summary
A personalized federated learning method performed by a processor of a user terminal operating in conjunction with a server, the method comprising: generating input data based on user data received through an interface of the user terminal; inputting the input data into a first learning model provided in the user terminal and training the first learning model using the corresponding output; transmitting local parameters for weights of a neural network included in the first learning model to the server; receiving global parameters derived based on the local parameters from the server; and inputting the input data into the first learning model, to which the global parameters are applied, and a second learning model associated with the first learning model, and training the second learning model using the corresponding output.