Patch-Based Neural Network Training With Noise for Image Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training neural networks in image recognition lack effective privacy protection, particularly when dealing with personal information, and often compromise recognition performance.
Innovation Solution
Divide input images into patches, process them through separate Upper and Lower models, adding noise to the Upper model's outputs, and integrate results for training, while storing patches across multiple servers to prevent data reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural network is divided into multiple models to process patches, then privacy protection is improved, but device complexity increases
Solution Approach 1:
The neural network is divided into multiple first models that process different patches of images independently. Each first model processes a specific patch without accessing other patches, enabling privacy protection while maintaining the overall functionality through the second model that integrates results.
Solution Approach 2:
A second model is introduced as an intermediary that receives calculation results from multiple first models and integrates them to produce final output. This intermediary structure enables privacy protection by preventing direct access to other patches while maintaining system functionality through coordinated processing.
2Reliability
If noise is added to calculation results, then privacy protection is improved, but measurement precision deteriorates
Solution Approach 1:
Noise is added to the calculation results of the first models as a privacy protection mechanism. The noise appears harmful to individual patch processing but serves a beneficial purpose by preventing identification of personal information while the second model integrates results to maintain overall recognition accuracy.
Data Source
AI summary
A learning model includes a plurality of first models and a second model different from the first models. An information processing method includes: a dividing step of dividing an image to be used for training of a learning model into a plurality of patches; a first input step of inputting the patches to a plurality of first models without overlapping; an adding step of adding noise to each of a plurality of calculation results output from the first models; and a second input step of inputting, to a second model, a plurality of calculation results to each of which the noise has been added.