Neural Data Representations for Differential Privacy Sharing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing privacy-preserving techniques for training AI models face challenges in ensuring data privacy, including re-identification, data leakage, bias propagation, model inversion attacks, and membership inference attacks, particularly in sensitive fields like healthcare, where personal information must be protected.
Innovation Solution
A privacy-preserving method using a trained neural network to generate representations of datasets that satisfy differential privacy by learning a noise model, ensuring private features are not retrievable while maintaining access to important information, achieved through a loss function with private and interest reward terms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If noise is added to model outputs to achieve differential privacy, then individual privacy is protected, but the quantity and quality of retrievable information decreases
Solution Approach 1:
The noise model is trained in advance on representative training datasets before being applied to protect the actual dataset. This preliminary training allows the system to learn the appropriate noise characteristics needed to achieve differential privacy while preserving information about non-private features, avoiding the need to add noise ad-hoc which would require事后 verification of privacy guarantees
Solution Approach 2:
The system changes the parameters of the noise distribution based on the specific private feature and dataset characteristics. By training the noise model on representative data, the system adapts the noise parameters to achieve optimal privacy protection while minimizing information loss about important non-private features
2Reliability
If context-dependent noise is added per model output, then differential privacy condition is satisfied, but the complexity and computational overhead increases
Solution Approach 1:
Instead of determining context-dependent noise parameters at the time of each model output, the system pre-trains a noise model on representative training datasets. This preliminary action captures the context-dependent characteristics in advance, simplifying the actual privacy-preserving step to applying a pre-computed noise distribution
Solution Approach 2:
The system uses training datasets that are representative copies or simulations of the actual data distribution. By training on these representative copies, the system learns the noise characteristics needed for the actual data without having to process the actual sensitive data repeatedly, reducing computational overhead
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention relates to a computer-implemented method for sharing a dataset while preserving privacy of the dataset, the dataset comprising a private feature F1, the method comprising the steps of: providing the dataset as input into a trained neural network, retrieving as output of the trained neural network, a plurality of representations of the dataset, selecting a representation among the plurality of representations, sharing the representation, characterized in that the trained neural network has been trained on training datasets representative of the distribution of the dataset, each training dataset comprising a training private feature corresponding to the private feature of the dataset, so that the private feature cannot be retrieved from any representation of the plurality of representations of the dataset. The invention also relates to a system for carrying out the above method and to a method for training a neural network.