Distributed Machine Learning Model Conversion System
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The widespread adoption of machine learning technology is hindered by concerns over data security and algorithm protection, as users are cautious about sharing sensitive information and providers are hesitant to disclose their models due to risks of unintended diversion and reverse engineering.
Innovation Solution
A conversion system that processes input data using a trained machine learning model, transmitting and receiving abstracted intermediate outputs rather than raw data, ensuring data security and protecting the model's integrity by not requiring the entire model to be shared with the client device.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If users provide raw data to third parties for machine learning, then machine learning processing can be performed, but data security and privacy protection deteriorate
Solution Approach 1:
The trained model is divided into multiple parts: an input-side part retained by the user and an output-side part held by the provider. Only intermediate outputs from the input-side part are transmitted, not raw data or the complete model, thereby enabling machine learning processing while protecting both data security and model integrity
Solution Approach 2:
Intermediate outputs serve as an intermediary representation between raw input data and final results. These abstracted intermediate values allow the model processing to occur without exposing sensitive raw data to third parties, thus maintaining data security while enabling versatile machine learning applications
2Productivity
If providers share the entire trained model with clients, then processing capability is improved, but model protection and security deteriorate
Solution Approach 1:
The trained model is segmented into an input-side conversion processing unit kept by the user and an output-side conversion processing unit kept by the provider. This segmentation enables the system to utilize the model's processing capability through distributed computation while preventing algorithm leakage by ensuring no single party possesses the complete model
Solution Approach 2:
Instead of providing the entire trained model to the client, only the necessary input-side portion is provided. The client performs partial processing locally and transmits only intermediate results to the provider for completion, thereby achieving sufficient processing capability while minimizing model exposure and protection risks
3Extent of automation
If complete models are distributed to client devices, then processing autonomy is improved, but device complexity and resource requirements worsen
Solution Approach 1:
The model is divided such that the client device retains only the input-side conversion processing unit, which is smaller and less complex than the complete model. This segmentation enables the client to perform autonomous initial processing without requiring the computational resources and storage capacity needed for the entire model
Data Source
Figure 1
Figure 2
Figure 3
AI summary
[Problem to be Solved] To provide a secure system capable of satisfying the requirements of both users and providers of machine learning technology. [Solution] Provided is a conversion system in which a client device includes: an input-side conversion processing unit that is a part of the trained model extending from an input layer to a first middle layer of the trained model, and performs conversion processing based on the input data to generate a first intermediate output of the first middle layer of the trained model; a client-side transmitting unit that transmits the first intermediate output to a server; a client-side receiving unit that receives a second intermediate output from the server, the second intermediate output being generated in the server based on the first intermediate output and being a conversion output of the second middle layer closer to the output side than the first middle layer of the trained model; and an output-side conversion processing unit that is a part of the trained model extending from the second middle layer to an output layer of the trained model, and performs conversion processing based on the second intermediate output to generate the output data.