Split Neural Network Inference for Privacy-Preserving Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models deployed on user-end devices face privacy concerns due to transmission of personal data to external servers, which may compromise user data and introduce additional computational overhead and latency.
Innovation Solution
A neural network model is split across multiple devices, with critical layers handling personal information on local devices and non-critical layers on remote servers, ensuring privacy preservation without degrading model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If ML models are deployed on external servers to execute machine learning tasks, then computational capabilities are improved, but user data privacy is compromised
Solution Approach 1:
The neural network model is divided into multiple segments (first NN sub-model and second NN sub-model) distributed across different devices. The first sub-model remains on the user-end device while the second sub-model is deployed on the external server, enabling collaborative inference that preserves privacy while maintaining computational capability.
Solution Approach 2:
Feature maps serve as an intermediary representation between the local and remote model components. The first NN sub-model processes personal data locally to generate feature maps, which are then transmitted to the second NN sub-model for further processing, eliminating the need to transmit raw personal data while achieving accurate predictions.
2Object-affected harmful factors
If ML models are encrypted or redesigned to ensure data privacy, then privacy protection is improved, but computational overhead and latency increase
Solution Approach 1:
By segmenting the neural network model across devices, the system avoids the need for complex encryption or redesign of the entire model. The local processing of personal data by the first NN sub-model eliminates encryption overhead, while only compressed feature maps are transmitted, reducing latency.
Solution Approach 2:
The critical processing of personal data is extracted and performed locally by the first NN sub-model, removing the need for encryption during transmission. Only the extracted feature maps, which do not contain personally identifiable information, are transmitted to the external server, eliminating privacy-related computational overhead.
3Measurement precision
If personal data is transmitted to external devices for processing, then model execution accuracy is improved, but data security is compromised
Solution Approach 1:
Personal data is transformed into feature maps through local processing by the first NN sub-model. These feature maps extract essential information while removing personally identifiable information, allowing accurate predictions to be made on external devices without compromising data security.
Solution Approach 2:
Feature maps act as an intermediary representation that bridges the need for accurate model execution and data security. The feature maps contain sufficient information for the second NN sub-model to make accurate predictions while not revealing sensitive personal information, thus maintaining both accuracy and security.
Data Source
AI summary
A first electronic device and a method for implementation for privacy-preserving splitting of neural network (NN) models for prediction across multiple devices is disclosed. The first electronic device applies an NN model on a received input datapoint to determine a set of feature maps. Each feature map is compared with the received input datapoint to determine a first NN sub-model including a first sub-set of NN layers. The determined first NN sub-model is applied on the received input datapoint to determine a first feature map. The first electronic device transmits the determined first feature map and information indicative of a second sub-set of NN layers of a second NN sub-model to a second electronic device. The second NN sub-model is used to determine a prediction output. The first electronic device receives the determined prediction output and controls rendering of the received prediction output on a display device.