Selective PII Encryption in Images for ML Usability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for protecting personally identifiable information (PII) in images and videos are inadequate, as they often result in unusable images for machine learning algorithms and are prone to security breaches due to storage overhead and inability to reverse the encryption process.
Innovation Solution
A cryptographic protection feature that uses a combinatory symmetric and asymmetric encryption technique to detect and encrypt PII in images and videos, allowing for reversible protection, enabling secure storage and use of images for machine learning without exposing raw data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional protection mechanisms are used to scrub images of PII, then security is improved, but the images become incapable of being used to train machine learning algorithms
Solution Approach 1:
The image is segmented into two distinct parts: encrypted regions containing PII and unencrypted regions containing useful data. This segmentation allows different portions of the image to serve different purposes - the encrypted portions protect sensitive information while the unencrypted portions remain available for machine learning analysis, thus resolving the contradiction between security and usability
Solution Approach 2:
Different quality levels are applied to different regions of the image. The PII-containing regions are encrypted to provide high security, while the non-PII regions remain unencrypted to maintain high usability. This local differentiation allows the image to simultaneously achieve both security and machine learning compatibility
2Reliability
If encryption is applied to protect PII, then security is improved, but storage overhead increases
Solution Approach 1:
The encryption overhead is extracted and isolated to only the necessary portions of the image (the PII regions), rather than encrypting the entire image. This extraction approach minimizes the amount of data that requires encryption resources, thereby reducing storage overhead while maintaining security for the critical PII portions
Solution Approach 2:
Instead of applying full encryption to the entire image, the solution applies encryption partially - only to the specific regions containing PII. This partial action is sufficient to achieve the security goal while avoiding the excessive storage overhead that would result from encrypting the complete image
3Reliability
If PII is made unreadable through encryption, then security is improved, but the ability to use the data for training machine learning algorithms is lost
Solution Approach 1:
The image data is segmented into encrypted PII regions and unencrypted useful data regions. This segmentation ensures that information loss is localized only to the PII portions that require protection, while the non-PII portions retain full readability and utility for machine learning training, thus preventing unnecessary loss of information
Solution Approach 2:
Different readability qualities are applied locally across the image - encrypted (unreadable) quality for PII regions and unencrypted (readable) quality for non-PII regions. This local quality differentiation ensures that security measures do not cause unnecessary loss of information in regions where PII is not present
Data Source
AI summary
Techniques for cryptographically protecting personally identifiable information in images and videos are described herein. An image may be obtained. One or more regions in the image may be detected based on an object detection algorithm. Pixels for each region of the one or more regions may be encrypted using a symmetric encryption technique and an initialization vector. The encrypted pixels for each region of the one or more regions may be written back into the image. A symmetric key of the symmetric encryption technique and the initialization vector may be encrypted using an asymmetric encryption technique. Metadata of the image may be updated to store the encrypted symmetric key and the encrypted initialization vector.