Federated Learning for Personalized Risk Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current risk prevention and control systems face challenges in protecting user privacy and reducing server computing pressure, as they require centralized data upload and use a single risk prevention and control model for all users, which lacks flexibility and fails to cater to individual user habits and preferences.
Innovation Solution
A personalized risk prevention and control system based on shared learning between terminal devices and servers, where initial risk prevention and control models are trained locally and fused across devices within classification groups, allowing for personalized models that protect user data and reduce server interaction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If centralized data upload to server is implemented for risk control decisions, then server computing capability can be utilized for analysis, but user privacy data security is compromised
Solution Approach 1:
The patent extracts the training data from the centralized server processing model and places it locally on terminal devices. Only model parameters and gradients are transmitted to the server, while the actual user data remains on-device, thus extracting the harmful data transmission aspect while retaining beneficial centralized coordination
Solution Approach 2:
The patent introduces federated learning as an intermediary mechanism between terminal devices and server. This intermediary enables collaborative model training without direct data transmission, using encrypted gradient exchanges as the mediator to achieve both privacy protection and server-based coordination
2Reliability
If terminal device performs centralized request to server, then risk control decisions can be made, but server resource consumption increases
Solution Approach 1:
The patent applies local quality by enabling terminal devices to perform local model inference using the trained risk control model. Each device independently evaluates its own data locally, eliminating the need for continuous centralized processing and significantly reducing server resource consumption while maintaining decision accuracy
3Device complexity
If same risk prevention and control model is used for all users, then model simplicity is maintained, but flexibility and personalization for different user habits and preferences deteriorates
Solution Approach 1:
The patent segments the user base into different classification groups based on user attributes and behaviors. Each group receives a customized risk control model trained on group-specific data, dividing the universal model into personalized segments that maintain simplicity within each segment while achieving overall flexibility and adaptability
Solution Approach 2:
The patent applies local quality by training different risk control models for different user classification groups. Each group receives a model tailored to its specific characteristics, habits, and preferences, making the system adaptable and versatile while maintaining model simplicity within each localized group
Data Source
AI summary
In an implementation, an initial risk prevention and control model of a to-be-trained target service corresponding to a classification group that a terminal device belongs to is received from a server. To obtain a risk prevention and control sub-model corresponding to the terminal device and based on pre-stored training sample data including at least data related to a user of the terminal device and the to-be-trained target service, performing model training on the initial risk prevention and control model. To obtain a risk prevention and control model corresponding to the classification group that the terminal device belongs to, sending to the server the risk prevention and control sub-model. The risk prevention and control model is received from the server as acquired data of the to-be-trained target service. Risk prevention and control processing is performed on the acquired data of the to-be-trained target service.