Distributed GAN Synthetic Data Sharing for Confidential ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of sharing sensitive data for machine learning tasks is hindered by trust issues between organizations, and existing methods like data obfuscation can lead to re-identification, compromising privacy.
Innovation Solution
A method using a distributed generative adversarial network (GAN) to generate synthetic training data, where the generator and discriminator functions are distributed across separate systems, allowing organizations to collaborate without revealing confidential data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data obfuscation is used to protect privacy, then data confidentiality is improved, but data leakage occurs allowing re-identification
Solution Approach 1:
The patent creates synthetic copies of real data records that preserve the statistical properties and patterns necessary for machine learning while eliminating personally identifiable information. The synthetic data is generated through a process that learns from real data and produces artificial records that are statistically equivalent but contain no actual sensitive information, thus preventing re-identification while maintaining data utility.
Solution Approach 2:
The patent introduces a synthetic data generation system as an intermediary between the data owner and the data processor. This intermediary creates synthetic training data that can be shared without revealing sensitive information, acting as a mediator that enables collaboration while protecting confidentiality. The synthetic data serves as a substitute that maintains the essential characteristics needed for model training without containing actual sensitive records.
2Measurement precision
If real sensitive data is shared for machine learning, then model training quality is improved, but organizational trust and privacy are compromised
Solution Approach 1:
The patent generates synthetic copies of sensitive data that retain the statistical properties, distributions, and relationships necessary for high-quality model training. These synthetic records are created to be indistinguishable from real data in terms of their utility for machine learning, while containing no actual sensitive information that could compromise privacy or trust.
Solution Approach 2:
The patent transforms real sensitive data into synthetic data by changing the fundamental parameter of data authenticity while preserving statistical properties. The transformation process modifies the data representation to eliminate personally identifiable information while maintaining the underlying patterns, distributions, and relationships that are essential for effective model training.
3Reliability
If synthetic data is generated from real data, then data confidentiality is improved, but data utility for machine learning must be maintained
Solution Approach 1:
The patent creates high-fidelity synthetic copies of real data that preserve the statistical properties, feature distributions, and relationships necessary for machine learning. The synthetic data is generated to be statistically equivalent to real data, ensuring that models trained on synthetic data achieve comparable performance to those trained on real data, thus maintaining data utility while protecting confidentiality.
Solution Approach 2:
The patent applies parameter transformations that change the authenticity of data while preserving its statistical characteristics. The transformation process modifies key parameters such as removing personally identifiable information while maintaining feature distributions, correlations, and patterns that are essential for machine learning utility.
Data Source
AI summary
A computer-implemented method of protecting confidentiality when generating synthetic training records for machine learning from sensitive data records, comprising a source computer system connected to a separate target computer system, where the target computer system comprises sensitive data records comprising private or confidential data. The source computer system performs the functions of a generator component of a generative adversarial network (GAN) and the target computer system performs the functions of a discriminator component of the GAN, where the generator and discriminator functions of the GAN are distributed between the source and target computer systems. Synthetic training records are generated using a computational process that does not reveal contents of the sensitive data records. Also disclosed is a method of training a machine-learning model using the one or more synthetic training records.