Generative Image Feature Replacement for Privacy-Preserving ML
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image anonymization methods, such as pixelation or blurring, reduce the accuracy of machine learning models, and existing synthetic methods are limited to facial features, failing to meet diverse regulatory requirements and maintaining data usability.
Innovation Solution
A method using generative image generation models to replace image features with synthetically created replacement features, employing instance segmentation and object detection models, and optionally reinforcement learning, to anonymize various features accurately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If obscuration methods (pixelation or blurring) are applied to anonymize image features, then privacy compliance is improved, but machine learning model accuracy deteriorates
Solution Approach 1:
The patent creates synthetic copies of image features (faces, objects, text) that replicate the structural and contextual information of original features while removing identifiable characteristics. These synthetic features are generated using generative models trained on annotated data, allowing the copies to maintain geometric relationships, spatial positioning, and contextual relevance needed for machine learning training while ensuring privacy compliance through non-identifiable content.
Solution Approach 2:
The patent transforms image features by changing their visual parameters (texture, color, detailed geometry) while preserving higher-level structural parameters (shape, spatial arrangement, contextual relationships). The generative models modify pixel-level characteristics to anonymize features while maintaining the essential information needed for training, effectively separating privacy protection from information preservation through parameter transformation.
2Reliability
If existing synthetic anonymization methods are used, then privacy compliance is improved, but adaptability to different feature types deteriorates
Solution Approach 1:
The patent implements a universal anonymization framework using generative image generation models that can handle multiple feature types (faces, objects, text, signs) within a single system. The model is trained on diverse annotated data representing various feature categories, enabling it to automatically adapt to different feature types without requiring separate anonymization methods for each category, thus achieving both privacy compliance and broad adaptability.
Solution Approach 2:
The patent employs dynamic generative models that can adapt their generation process based on the input feature type and context. The system dynamically adjusts the anonymization approach by analyzing the characteristics of each feature and selecting appropriate generation parameters, allowing flexible handling of diverse feature types while maintaining consistent privacy protection standards.
3Ease of manufacture
If obscuration methods are applied to anonymize image features, then processing simplicity is improved, but data usability for machine learning deteriorates
Solution Approach 1:
The patent creates synthetic copies that preserve the functional information needed for machine learning while removing identifying characteristics. These copies maintain the geometric relationships, spatial positioning, and contextual information that make the data useful for training models, effectively replacing the need for complex obscuration methods while improving data usability through generation-based synthesis.
Data Source
Figure 1
Figure 2a~2c
Figure 3a~3b
AI summary
The present invention is concerned with a method, in particular a computer-implemented method, for anonymizing features (F) in an image, the method comprising the steps of: - receiving at least one image (I(F)) comprising at least one feature (F); - identifying the at least one feature (F) and detecting feature boundaries of the at least one feature (F); and - replacing the at least one feature (F) by a replacement feature (RF) generated by using a generative image generation model (IGM), and - outputting an anonymized image (IA) comprising the replacement feature (RF). The invention is also concerned with a preferred use of the method for training a machine learning model, a computer program, and a computer-readable (storage) medium.