Distributed Neural Network Layout to Prevent Model Imitation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional neural network configurations in image capturing devices with low computational power are vulnerable to imitation due to the transmission of computation results to servers, which can be intercepted and used to replicate the network configuration.

Innovation Solution

Distribute intermediate layers of the neural network across an image capturing apparatus and a server with higher computational power, ensuring the input and output layers are on the image capturing apparatus, and prevent data interception by transmitting and processing intermediate data within a secure communication network.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If intermediate layers and output layer are disposed on a server with high computational power, then processing time is reduced, but the neural network configuration becomes vulnerable to imitation due to data transmission

Engineering Contradiction:

Improveprocessing timeVSAvoidsecurity of neural network configuration

Core Design Contradiction:

Loss of timeVSReliability

Solution Approach 1:

The neural network is segmented into multiple layers distributed across different devices. Specifically, the input layer and output layer are disposed on the image capturing apparatus, while intermediate layers are disposed on both the image capturing apparatus and the server. This segmentation allows processing to be distributed, reducing the processing time while preventing complete imitation of the neural network configuration since not all layers are accessible to any single device.

Inventive Principle:

Principle #1Segmentation

2Measurement precision

If more intermediate layers are added to the neural network, then reasoning accuracy is improved, but computational load increases and processing time increases

Engineering Contradiction:

Improvereasoning accuracyVSAvoidprocessing time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The intermediate layers are segmented and distributed between the image capturing apparatus and the server. The server, with its higher computational power, handles a portion of the intermediate layer processing, while the image capturing apparatus handles other portions. This distribution enables the system to maintain a large number of intermediate layers for high reasoning accuracy while reducing the processing time burden on any single device.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system merges the computational resources of the image capturing apparatus and the server to handle the intermediate layer processing. By combining the processing capabilities of both devices, the system can process more intermediate layers in parallel, thereby maintaining high reasoning accuracy while reducing overall processing time.

Inventive Principle:

Principle #5Merging (Combining)

Data Source

PatentUS12541965B2Information processing system capable of preventing imitation of configuration of neural network, method of controlling image processing system, and storage medium

Publication Date: 2026.02.03 CANON KK

AI summary

An information processing system capable of preventing imitation of the configuration of a neural network. The information processing system includes an image capturing apparatus and a server that are communicable with each other. In the information processing system, an input layer, a plurality of intermediate layers, and an output layer form the neural network. The input layer and the output layer are disposed in one of the image capturing apparatus and the server. Further, the intermediate layers are disposed such that the intermediate layers are distributed between the first apparatus and the second apparatus.