Secure Deep-Learning Model Service via Client-Side Network Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep-learning model services face challenges in providing secure domain-specific tuning and data protection, as existing methods like homomorphic encryption face issues with noise accumulation and decreased accuracy, while federated learning struggles with global optimization and model contamination risks.
Innovation Solution
A secure online deep-learning model service is proposed, where clients generate user-private input and output networks by fine-tuning the deep-learning model using user data, and only the hidden network is managed by the server, ensuring user data privacy and preventing data reconstruction or interpretation by the service company.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If homomorphic encryption is used to protect user data, then data security is improved, but computation speed decreases and accuracy deteriorates due to noise accumulation
Solution Approach 1:
The deep learning model is segmented into three parts: input network (client-side), hidden network (server-side), and output network (client-side). This segmentation allows homomorphic encryption to be applied only where necessary for security while avoiding its computational overhead in the main processing path, thus resolving the contradiction between data security and computation speed.
Solution Approach 2:
The sensitive input and output processing is extracted from the server-side hidden network and moved to client-side networks. This extraction eliminates the need for homomorphic encryption during the core computation phase, maintaining computation speed while preserving data security through encrypted input/output handling.
2Reliability
If homomorphic encryption is used for deep-learning operations, then data security is improved, but manufacturing precision decreases due to operation estimation in nonlinear networks
Solution Approach 1:
By segmenting the model into input network, hidden network, and output network, homomorphic encryption is applied only to the input and output stages rather than throughout the entire model. This limited application maintains data security while avoiding accuracy degradation that would result from encrypting all nonlinear operations.
Solution Approach 2:
The client-side input and output networks act as intermediaries that handle encrypted data transformation before and after the server-side processing. This intermediary approach allows secure data handling without requiring homomorphic encryption during the critical hidden layer computations, preserving model accuracy.
3Reliability
If federated learning is used to ensure security, then data privacy is improved, but model optimization deteriorates due to difficulty in creating global optimization models
Solution Approach 1:
The model is segmented such that the hidden network (containing the core learned patterns) is trained centrally on the server with full data access, while input and output networks are customized locally. This segmentation enables global optimization of the hidden network while maintaining data privacy, resolving the contradiction between privacy and optimization.
Solution Approach 2:
The critical optimization-capable hidden network is extracted from local devices and centralized on the server, allowing it to be trained on aggregated data from multiple users. This extraction enables global model optimization while keeping raw data localized, achieving both privacy and optimization goals.
4Reliability
If federated learning is used for domain-specific tuning, then data security is improved, but device complexity increases as local devices are burdened with entire model operations
Solution Approach 1:
The model is segmented into lightweight input and output networks that reside on local devices, and a hidden network that resides on the server. This segmentation significantly reduces the computational burden on local devices while maintaining data security, as only the smaller input/output networks need to be processed locally.
Solution Approach 2:
The computationally intensive hidden network is extracted from local devices and placed on the server, relieving local devices of the burden of processing large model parameters. This extraction reduces device complexity while maintaining security through the client-side input/output networks.
Data Source
AI summary
Disclosed herein are an apparatus and method for a secure online deep-learning model service. A client apparatus for the secure online deep-learning model service includes memory in which a secure input network and a secure output network generated in advance based on a deep-learning model provided from a server and at least one program are recorded and a processor for executing the program. The program may perform generating an input vector by passing user data through the secure input network, transmitting the input vector to the server, receiving an output vector acquired by passing the input vector through the hidden network of the deep-learning model from the server, and generating output data by passing the received output vector through the secure output network.