ML Image Encoding for Selective Object Privacy Protection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image encoding technologies fail to effectively alter specific objects in images for privacy while preserving the recognizability of other objects, and often require manual identification and labeling of objects for privacy.
Innovation Solution
Utilizing machine learning models generated in conjunction with object-recognition models to encode images, altering specific objects like faces or license plates for privacy while maintaining the recognizability of other objects, and optimizing the encoding process through iterative training to minimize recognition errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are used to encode images for privacy protection, then specific objects become unrecognizable to object-recognition models, but the complexity of the encoding system increases
Solution Approach 1:
The patent introduces an intermediary encoding model that acts as a mediator between the original image and the object-recognition model. This encoding model transforms the image in a way that preserves privacy by making specific objects unrecognizable, while still allowing the system to function. The intermediary encoding layer effectively decouples the privacy protection function from the object recognition function, resolving the contradiction between privacy effectiveness and system complexity.
Solution Approach 2:
The patent applies parameter changes by modifying the image representation through the encoding model's parameters and weights. By adjusting and optimizing these parameters during training, the system achieves effective privacy protection without requiring fundamental changes to the overall system architecture. The parameter optimization allows the same system structure to adapt to different privacy protection requirements.
2Measurement precision
If iterative training is used to optimize encoding models, then recognition errors are minimized, but the training time and computational resources increase
Solution Approach 1:
The patent implements feedback mechanisms during the iterative training process, where the encoding model receives feedback from the object-recognition model's performance on encoded images. This feedback loop allows the encoding model to continuously improve its ability to protect privacy while maintaining recognition accuracy for non-protected objects. The feedback-driven optimization achieves high precision without requiring excessive training iterations by focusing improvements where they are most needed.
Solution Approach 2:
The patent applies partial action by focusing the iterative training optimization on specific aspects of the encoding model that have the greatest impact on privacy protection and recognition accuracy. Rather than uniformly optimizing all parameters throughout extensive training, the system identifies and optimizes the most critical parameters, achieving satisfactory results with reduced training time and computational resources.
3Reliability
If all objects in images are encoded for privacy, then privacy protection is maximized, but the recognizability of important objects is lost
Solution Approach 1:
The patent applies local quality by making different parts of the image undergo different transformations. Specifically, the encoding model is trained to apply strong privacy protection transformations to sensitive objects (like faces and license plates) while applying minimal or no transformations to non-sensitive objects. This selective approach ensures that privacy protection is applied locally where needed, preserving the recognizability of important objects that do not require privacy protection.
Solution Approach 2:
The patent segments the image processing task by differentiating between sensitive objects that require privacy protection and non-sensitive objects that should remain recognizable. The encoding model learns to segment the image content and apply appropriate encoding strategies to different segments, thereby maximizing privacy protection coverage for sensitive objects while maintaining information integrity for important objects.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure is directed to encoding images. In particular, one or more computing devices can receive data representing one or more machine learning (ML) models configured, at least in part, to encode images comprising objects of a particular type. The computing device(s) can receive data representing an image comprising one or more objects of the particular type. The computing device(s) can generate, based at least in part on the data representing the image and the data representing the ML model(s), data representing an encoded version of the image that alters at least a portion of the image comprising the object(s) such that when the encoded version of the image is decoded, the object(s) are unrecognizable as being of the particular type by one or more object-recognition ML models based at least in part upon which the ML model(s) configured to encode the images were trained.