Privacy-Preserving Data Transformation in Machine Learning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI/ML systems face ethical and regulatory challenges due to the lack of theoretical guarantees for protecting individual privacy when using large datasets containing sensitive information, leading to the withholding of valuable data.
Innovation Solution
A method involving a machine learning model with a data transformation module, sensitive attribute suppression module, annotated useful attribute preservation module, and unannotated useful attribute preservation module, which uses neural networks to transform data while ensuring sensitive attributes are suppressed and useful attributes are preserved, with a total loss function for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If organizations use large datasets containing sensitive data in AI/ML operations, then model generalization and utility are improved, but ethical and regulatory concerns regarding privacy leakage worsen
Solution Approach 1:
The patent segments the dataset into sensitive attributes and non-sensitive attributes, applying different processing strategies to each. The sensitive attributes are suppressed through dedicated suppression modules while non-sensitive attributes are preserved for model training, allowing the system to utilize large datasets without compromising privacy.
Solution Approach 2:
The patent introduces an intermediary data transformation process that acts as a mediator between the original sensitive data and the AI/ML model. This transformation module suppresses sensitive attributes while preserving useful information, enabling model training without direct exposure to sensitive data.
2Reliability
If organizations obfuscate sensitive data empirically, then privacy protection may be achieved, but theoretical guarantees for privacy protection are lost
Solution Approach 1:
The patent implements feedback mechanisms through loss functions that continuously measure and adjust the suppression effectiveness. The suppression loss, utility loss, and total loss provide continuous feedback to optimize the suppression process, ensuring both privacy protection and utility preservation are maintained throughout training.
Solution Approach 2:
The patent dynamically adjusts suppression parameters and transformation parameters during the training process based on loss function feedback. This allows the system to adaptively optimize the balance between privacy protection and data utility, providing both empirical and theoretical guarantees.
3Reliability
If organizations withhold valuable datasets to avoid privacy concerns, then ethical and regulatory compliance is improved, but data utility and model performance worsen
Solution Approach 1:
The patent extracts and removes sensitive attributes from the dataset while retaining the non-sensitive useful attributes. This extraction process allows organizations to release and utilize valuable datasets for AI/ML operations while eliminating the privacy risks associated with sensitive information.
Solution Approach 2:
The patent applies different quality standards to different parts of the data: sensitive attributes receive strong suppression to ensure privacy compliance, while non-sensitive attributes are preserved with high fidelity to maintain data utility. This localized quality approach allows simultaneous achievement of ethical compliance and data productivity.
4Object-affected harmful factors
If traditional data suppression methods are used, then sensitive data protection may be achieved, but useful attribute preservation and theoretical proofs are lacking
Solution Approach 1:
The patent employs dynamic suppression that adapts during the training process rather than applying static suppression rules. The suppression strength and transformation parameters are continuously adjusted based on loss function feedback, allowing the system to dynamically balance sensitive data protection with useful attribute preservation.
Solution Approach 2:
The patent creates a composite loss function that combines multiple objectives: suppression loss for sensitive data protection, utility loss for useful attribute preservation, and total loss for overall optimization. This composite approach integrates multiple competing requirements into a unified training framework.
Data Source
AI summary
In some aspects, the techniques described herein relate to a method including: executing a machine learning model; providing a data transformation module of the machine learning model that outputs a transformed dataset; providing a sensitive attribute suppression module of the machine learning model that outputs a sensitive attribute suppression loss; providing an annotated useful attribute preservation module of the machine learning model that outputs an annotated useful attribute preservation loss; providing an unannotated useful attribute preservation module of the machine learning model that outputs an unannotated useful attribute preservation loss; combining the sensitive attribute suppression loss, the annotated useful attribute preservation loss, and the unannotated useful attribute preservation loss into a total loss; and training a neural network of the data transformation module and a neural network of the unannotated useful attribute preservation module using the total loss.