AI Model Refinement via Gradient Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The refinement of artificial intelligence (AI) models in electronic devices poses a risk of sensitive user information leakage during transmission, particularly when context information including personal, biometric, and financial data is shared, highlighting the need for secure and efficient model updating mechanisms.
Innovation Solution
An electronic apparatus and server system that detects changes in context information, determines a gradient for refining a local AI model, transmits this gradient to a server, and receives updates from a globally refined model to iteratively refine the local model, thereby minimizing the transmission of sensitive data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If context information including personal, biometric, and financial data is transmitted to the server for AI model refinement, then the model refinement accuracy and performance are improved, but the risk of sensitive user information leakage increases
Solution Approach 1:
The patent extracts only the necessary gradient information from the context data for model refinement, leaving sensitive personal, biometric, and financial information behind in the electronic apparatus. This selective extraction allows the server to improve model accuracy using only the mathematical gradients needed for optimization, while the harmful sensitive data never leaves the user's device.
Solution Approach 2:
The patent introduces gradient information as an intermediary between the context data and the model refinement process. Instead of directly transmitting sensitive context information to the server, the electronic apparatus first processes this data locally to generate gradient representations, which serve as a safe mediator that conveys only the essential optimization signals without exposing underlying sensitive information.
2Productivity
If all context information is transmitted to the server for comprehensive model refinement, then the global model performance is improved, but the transmission data volume and processing time increase
Solution Approach 1:
The patent extracts only the gradient information necessary for model optimization from the comprehensive context data, dramatically reducing the data volume that needs to be transmitted to the server. This selective extraction maintains the efficiency benefits of comprehensive data processing while eliminating the overhead of transmitting and processing unnecessary sensitive information.
3Adaptability or versatility
If context information is transmitted to the server, then the AI model can be refined based on real-time data, but the transmission process creates security vulnerabilities
Solution Approach 1:
The patent uses gradient information as an intermediary that enables real-time model adaptation while protecting security. The gradients convey the essential adaptive signals from real-time context changes to the server for global model updates, but they do so through a mathematically transformed representation that does not expose the underlying sensitive context information during transmission.
Solution Approach 2:
The patent extracts only the adaptive gradient signals needed for real-time model refinement, separating these essential adaptation signals from the sensitive context information. This allows the system to maintain real-time adaptability while the harmful sensitive data remains securely localized in the electronic apparatus.
Data Source
AI summary
Provided are an artificial intelligence (AI) system simulating a function of a human brain, such as cognition and judgment, using a machine learning algorithm, such as deep learning, and an application thereof. Also, provided is a method, performed by an electronic apparatus, of refining an artificial intelligence (AI) model, the method including: detecting information about a context of an electronic apparatus used to refine a local model stored in the electronic apparatus being changed; determining a gradient for refining the local model based on the changed information about the context; refining the local model based on the determined gradient; transmitting the gradient to a server; receiving, from the server, information about a global model refined based on the gradient; and refining the local model based on the received information.