Facial Data Anonymization via Synthetic Image Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in processing visual data for innovative applications like autonomous driving and smart city analytics due to privacy requirements, which limit the use of original data and conventional anonymization methods are often incompatible with AI/ML algorithms.
Innovation Solution
A personal data processing system that includes an edge device capable of detecting facial image data and generating a descriptor with mimic expression and identification feature characteristics, which is then used to create an obfuscated image. This system splits the workload with a more powerful server device to generate artificial facial image data, ensuring seamless integration and compliance with privacy regulations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional anonymization methods (blurring, masking) are used, then privacy requirements are met, but compatibility with AI/ML algorithms is lost
Solution Approach 1:
The patent creates synthetic facial images that copy the essential characteristics of real faces (mimic expression features, anatomical structure) while replacing actual personal identification data. This allows AI/ML algorithms to process and train on facial data that appears real but contains no identifiable information about specific individuals, thus maintaining both privacy compliance and algorithm compatibility.
Solution Approach 2:
The system transforms facial images by modifying specific parameters related to identification features while preserving parameters related to mimic expression and anatomical structure. This selective parameter transformation enables the data to remain useful for AI training while removing identifiable characteristics, resolving the contradiction between privacy and usability.
2Reliability
If deep natural anonymization is used, then privacy requirements are met and natural appearance is maintained, but computational resources required increase significantly
Solution Approach 1:
The patent divides the anonymization process into separate functional modules: one for detecting facial features, another for generating synthetic faces based on preserved characteristics, and a third for integrating these elements. This segmentation allows each component to be optimized independently and reduces overall computational burden compared to monolithic deep anonymization systems.
Solution Approach 2:
The system performs preliminary detection and extraction of essential facial characteristics (mimic expression features, anatomical structure) before generating synthetic images. By preparing and storing these feature descriptors in advance, the system avoids repeated heavy computational analysis during the anonymization process, reducing energy consumption while maintaining quality.
3Reliability
If fully synthetic images are generated, then privacy requirements are met, but variety and representation of real life scenarios are insufficient
Solution Approach 1:
The patent uses detected facial characteristics and anatomical structure as intermediary elements that bridge real and synthetic data. By preserving mimic expression features and anatomical relationships from real faces while generating synthetic facial images, the system maintains the variety and realism of actual human faces while ensuring privacy compliance, thus avoiding the limitations of purely synthetic data.
Data Source
AI summary
Personal data processing system comprising a first device configured to capture image data including facial image data of a human being; the first device further configured to detect the facial image data of a human being, to determine a descriptor including information describing a characteristic of a mimic expression feature and a characteristic of each identification feature of a selectable subset of the plurality of identification features; the first device further configured to provide an obfuscated image with placeholder data replacing the facial image data; the first device further configured to provide the obfuscated image and the descriptor to a second device; the second device configured to perform a replacement of the placeholder data with artificial facial image data corresponding to the descriptor to produce an anonymized image.