Face Anonymization via Machine Learning Perturbation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional methods of facial anonymization in images, such as using white blobs or blurring, negatively impact the extraction of important information and may not comply with privacy regulations like GDPR, especially when recordings are stored for long periods.
Innovation Solution
A system and method that uses a machine-learning based face generator to produce a perturbed face with attributes perceived as substantially equivalent to the original, while being unrecognizable as the same person, by training a model with labeled images and feedback from viewing entities, including attribute classifiers and human viewers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional anonymization methods (white blobs, blurring, pixelation) are used, then privacy protection is improved, but information extraction effectiveness deteriorates
Solution Approach 1:
The patent applies parameter changes by transforming the face image through controlled modifications of facial parameters (such as altering facial features, expressions, or characteristics) while preserving the underlying identity and attribute information. This allows the face to remain recognizable as belonging to the same person while preventing direct identification, thus maintaining information extraction effectiveness while achieving privacy protection.
Solution Approach 2:
The patent uses copying by creating a synthetic or generated face image that replicates the structural and attribute characteristics of the original face but with modified visual details. The generated face preserves essential features for attribute extraction (such as age, gender, emotion) while altering sufficient visual characteristics to prevent direct recognition of the original person, effectively copying the functional properties while changing the identifying properties.
2Reliability
If traditional anonymization methods are used, then compliance with privacy regulations is improved, but the ability to extract valuable facial attributes deteriorates
Solution Approach 1:
The patent applies parameter changes by systematically modifying specific facial parameters while preserving others. The generated face maintains attribute-related parameters (such as age group, gender, emotional state) while changing identification parameters (specific facial features, unique characteristics). This selective parameter transformation enables compliance with privacy regulations that require anonymization while preserving the capability to extract valuable facial attributes for analysis.
Solution Approach 2:
The patent uses segmentation by separating the facial image into distinct functional components: identification features and attribute features. The anonymization process selectively modifies identification features while preserving attribute features. This segmentation allows independent manipulation of different facial characteristics, enabling compliance with privacy regulations while maintaining the ability to extract valuable attributes such as age, gender, and emotion.
3Reliability
If faces are replaced with white blobs or blurred, then personal identifiability is improved, but the visual quality and detail of the image deteriorates
Solution Approach 1:
The patent uses copying by generating a synthetic face image that replicates the visual quality, detail, and structural characteristics of the original face. The generated face maintains realistic appearance, proper lighting, texture, and anatomical accuracy, thereby preserving visual quality and detail. At the same time, the generated face is sufficiently different from the original to prevent personal identifiability, thus resolving the contradiction between visual quality and privacy protection.
Solution Approach 2:
The patent applies parameter changes by transforming visual parameters such as color, texture, feature shapes, and facial characteristics while maintaining the overall image quality. The generated face uses controlled parameter modifications to alter identifiability without degrading visual quality. Techniques such as preserving lighting conditions, maintaining anatomical proportions, and keeping texture realism ensure that the image detail remains high while personal identifiability is prevented.
Data Source
AI summary
A system and method of anonymization of a face in a set of images by at least one processor, the method including: receiving a first set of images; extracting from the first set of images a first face, depicting a person and having a first set of attributes; and performing a perturbation of the first face to produce a second face having a second set of attributes, wherein the second set of attributes is adapted to be visually perceived by a viewing entity as being substantially equivalent to the first set of attributes, and wherein the second face is adapted to be visually perceived by a viewing entity as not pertaining to the depicted person.