Patch-Based Image Recognition Learning for Privacy Protection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network learning techniques face challenges in achieving sufficient privacy protection, especially when handling sensitive information, as they either require large data communication or complicate the learning process, and cannot adjust data effectively for improved performance without risking privacy exposure.
Innovation Solution
The method involves segmenting images into small patches, distributing them across multiple servers, and using a two-part learning model configuration with upper and lower models to process and combine outputs, ensuring privacy by preventing restoration of the original image from a single location.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural networks are segmented to provide interim learning results to respective neural networks, then privacy protection can be considered, but original data can be restored from data in the course of computation, so sufficient privacy protection cannot be achieved
Solution Approach 1:
The patent divides the image data into multiple patches and processes each patch through separate neural networks. The segmentation is designed so that each neural network processes only a portion of the data, and the intermediate results cannot be used to restore the original image. This resolves the contradiction by ensuring that while segmentation enables distributed processing, the privacy protection is maintained because the intermediate results are insufficient for data restoration.
Solution Approach 2:
The patent extracts only the necessary features from each image patch for processing, rather than processing the complete original image. By taking out and processing only relevant portions through specialized neural networks, the system achieves privacy protection because the extracted features cannot be used to reconstruct the original sensitive data.
2Reliability
If image data is divided into patches and processed by respective neural networks, then privacy protection is improved, but the learning process becomes more complex
Solution Approach 1:
The patent designs neural networks with universal processing capabilities that can handle multiple types of image patches and features. By creating multi-functional neural network components that can process different patch types through standardized operations, the system reduces the overall complexity despite the segmented architecture, as the same processing framework applies across all patches.
Solution Approach 2:
The patent assigns different specialized neural networks to different image patches based on their specific characteristics and requirements. Each neural network is optimized for its local processing needs, which simplifies the overall system design by allowing specialized processing rather than requiring a single complex universal processor to handle all patches equally.
3Reliability
If original image is not centrally collected for learning, then privacy protection is enhanced, but the ability to adjust data for improved performance is limited
Solution Approach 1:
The patent introduces intermediate processing layers and feature extraction mechanisms that act as mediators between the distributed image patches and the learning process. These intermediaries enable data adjustment and performance optimization by transforming and processing the distributed data in ways that improve learning effectiveness while maintaining privacy, as the adjustments are made on processed features rather than original sensitive data.
Data Source
AI summary
An information processing method generates a learned model for image recognition, wherein a learning model includes a plurality of first models and a second model that is different from the first models. The information processing method includes: dividing an image that is to be used in learning into a plurality of patches; inputting each of the divided plurality of patches into a respective model of the plurality of first models, the respective model being predetermined for each of the patches, and performing calculation; and combining an output of a calculation result of each of the plurality of first models in the second model and generating a learned model by learning the learning model.