Adversarial Feature Elimination for Privacy-Preserving Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for eliminating sensitive information in machine learning models, such as data encryption and attribute elimination, often compromise the predictive power of the model, especially when attributes related to vulnerable populations like dementia are removed.
Innovation Solution
A machine learning-based method using adversarial networks to eliminate sensitive information by minimizing a loss function, ensuring the model can perform tasks without recognizing attributes like dementia while maintaining data utility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If attribute elimination is used to remove sensitive information, then privacy protection is improved, but predictive power of the model deteriorates
Solution Approach 1:
The patent extracts and eliminates only the specific sensitive attributes (e.g., dementia-related features) from the input feature set while retaining all other features that contribute to predictive power. This selective extraction approach protects privacy without compromising the model's ability to perform its primary task.
Solution Approach 2:
The patent segments the feature elimination process into two distinct components: (1) an information elimination model that removes sensitive information, and (2) an adversarial network that ensures task performance is maintained. This segmentation allows each component to optimize for its specific goal without interfering with the other.
2Object-affected harmful factors
If data encryption is used to protect sensitive information, then privacy protection is improved, but data processing capability deteriorates
Solution Approach 1:
Instead of encrypting data to protect it (which makes processing difficult), the patent inverts the approach by directly eliminating sensitive information from the feature space. This allows the model to process non-sensitive data freely while the sensitive information is simply not present, avoiding the processing limitations of encryption.
3Object-affected harmful factors
If sensitive attributes are removed to protect privacy, then privacy protection is improved, but data utility deteriorates
Solution Approach 1:
The patent selectively extracts and removes only the specific sensitive attributes related to vulnerable populations while preserving all other useful information in the dataset. This targeted extraction ensures that data utility is maintained for legitimate purposes while protecting specific sensitive information.
Solution Approach 2:
The adversarial network provides feedback to ensure that the elimination of sensitive information does not inadvertently remove useful information needed for task performance. The feedback mechanism monitors and maintains data utility while achieving privacy protection goals.
Data Source
AI summary
According to a method provided in the disclosure, based on an original signal and an information elimination (IE) model, a feature not including information allowing an attribute to be recognizable is generated. A task is then performed using a machine learning model based on the generated feature. For training the IE model, two adversarial networks are provided and a loss function is minimized. Input layers of the two adversarial networks are generated based on output layer and input features of the IE model. Generator of one adversarial network and discriminator of the other adversarial network are configured to perform the task, while discriminator of the one adversarial network and generator of the other adversarial network are configured to recognize the attribute. The loss function is associated with a disentangling loss of input layers of the two adversarial networks, as well as losses of each generator and discriminator.