Federated Dental Image Classification With Encrypted Model Sharing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
There is a lack of reliable auxiliary classification methods for tooth diseases in dental diagnosis, especially in areas with limited medical resources and skilled clinicians, and existing methods do not adequately protect patient privacy.
Innovation Solution
A tooth disease classification method based on federated learning (FL) that involves acquiring oral information, preprocessing dental images, training local models, encrypting model parameters, and performing fusion and classification across local servers to ensure accurate diagnosis without central data processing or privacy disclosure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If centralized data processing is used for tooth disease classification, then diagnostic accuracy can be improved through big data analysis, but patient privacy security deteriorates due to central storage of sensitive oral information
Solution Approach 1:
The patent divides the centralized classification system into multiple local servers distributed across different regions. Each local server independently processes and classifies dental images using locally trained models, segmenting the data processing function while maintaining collaborative learning through model parameter sharing. This segmentation eliminates the need to centralize sensitive patient data, thereby preserving privacy security while achieving accurate classification through distributed big data analysis.
Solution Approach 2:
The patent introduces model parameters as an intermediary between local data and the classification system. Instead of directly sharing or centralizing sensitive patient data, the system uses encrypted model parameters that capture diagnostic knowledge. These parameters act as a mediator that enables collaborative learning and accurate classification without exposing underlying patient information, thus resolving the contradiction between diagnostic accuracy and privacy security.
2Measurement precision
If data is transmitted to cloud servers for processing, then classification accuracy can be improved through access to more data, but data transmission security deteriorates due to potential exposure during transfer
Solution Approach 1:
The patent segments the data processing workflow so that sensitive patient data remains localized and is never transmitted to cloud servers. Only non-sensitive model parameters are communicated between local servers and the cloud platform. This segmentation maintains classification accuracy by enabling collaborative learning while eliminating data transmission security risks associated with sending sensitive patient information over networks.
Solution Approach 2:
The patent uses model parameters as copies that represent the essential diagnostic knowledge extracted from local data. Instead of transmitting original sensitive patient data to the cloud, the system transmits these parameter copies that capture the learned patterns. This copying approach enables accurate classification through cloud-based model updates while ensuring that no sensitive patient information is exposed during transmission or storage.
3Object-affected harmful factors
If local servers process data independently without collaboration, then data security is maintained, but diagnostic accuracy deteriorates due to limited local data samples
Solution Approach 1:
The patent merges the capabilities of multiple independent local servers through federated learning. Each server maintains its data security by processing locally, but they combine their diagnostic knowledge by sharing and aggregating model parameters. This merging enables the system to achieve high classification accuracy equivalent to centralized processing while each server independently maintains data security, as the collaboration occurs at the model parameter level rather than the data level.
Data Source
AI summary
The present disclosure provides a tooth disease classification method based on federated learning (FL), including the following steps: S1: acquiring oral information of a user; S2: transmitting an oral dental image to a local server for preprocessing; S3: performing, by the single local server, model training according to a preprocessed oral dental image set and a model of a cloud master server; S4: transmitting trained model parameters to the cloud master server through digital encryption; and S5: performing, by each of local servers, fusion and classification according to digitally decrypted model parameters from the cloud master server. The method provided by the present disclosure has accurate classification, does not need to centrally process data, and keeps the oral information strictly secret without privacy disclosure; and meanwhile, the method encrypts the model parameters to further ensure the security.