Image Concealment And Region Labeling For Accurate Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for machine learning using personal information concealed images reduce accuracy due to the use of unnatural images, and the use of original images for training is hindered by privacy concerns.
Innovation Solution
An information processing apparatus that determines and conceals personal information regions in images, applies label information to these regions, and displays the labeled concealed images for training, allowing the use of original images while protecting privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If personal information concealed images are used in machine learning training, then privacy protection is improved, but training accuracy deteriorates due to unnatural image quality and loss of original features
Solution Approach 1:
The image data is segmented into two distinct parts: the original image data used for training and the concealed image data used for display. This segmentation allows the training process to access full-quality original images while the display shows privacy-protected concealed images, thereby resolving the contradiction between privacy protection and training accuracy
Solution Approach 2:
Label information serves as an intermediary element that connects the concealed image display with the original image training. The labels applied to concealed regions provide semantic information that bridges the gap between the privacy-protected display and the accuracy-requiring training process
2Measurement precision
If original images are used for machine learning training, then training accuracy is improved, but privacy protection deteriorates as personal information becomes exposed
Solution Approach 1:
The system segments the image handling process into distinct training and display pathways. Original images are routed to the training pathway where accuracy is critical, while concealed images are routed to the display pathway where privacy protection is critical. This segmentation eliminates the need to choose between accuracy and privacy
Solution Approach 2:
Different quality levels are applied to different parts of the system: full-quality original images are used locally in the training process where accuracy is needed, while privacy-protected concealed images are used locally in the display process where privacy is needed. Each part receives the appropriate quality level for its specific function
3Reliability
If concealed regions are applied to protect personal information, then privacy protection is improved, but information utility deteriorates as training models cannot learn from obscured areas
Solution Approach 1:
The information flow is segmented into two parallel channels: one channel preserves the complete original image information for training purposes, while the other channel applies concealment for display purposes. This segmentation ensures that information utility is maintained in the training channel while privacy protection is achieved in the display channel
Solution Approach 2:
The system creates a copy of the image data for training purposes while using the concealed version for display. The original uncopied image data retains full information utility for training, while the copied concealed data provides privacy protection for display, eliminating the information loss problem
Data Source
AI summary
An information processing apparatus comprising: a determining unit configured to determine a concealed region for concealing personal information in a first image; a concealing unit configured to execute concealing processing on the first image on a basis of the concealed region; an applying unit configured to apply label information to the concealed region; a display unit configured to display the label information superimposed on a second image, which is a concealed image obtained by executing the concealing processing; and a training unit configured to perform training using the first image.