Federated Synthetic Data Generation for Privacy-Preserving ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Financial institutions face challenges in using their transaction data for machine learning model training due to privacy protection laws, which prevent the use of personal identifiable information.
Innovation Solution
The implementation of federated, collaborative, privacy-preserving learning models generates synthetic data using local and global synthetic data generating models, such as generative adversarial networks, to create anonymized data for training machine learning models, including fraud detection and intelligent services like anomaly detection and payment trend prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real transaction data containing PII is used for machine learning model training, then model training quality is improved, but privacy protection requirements are violated
Solution Approach 1:
The patent creates synthetic copies of real transaction data that preserve statistical properties and patterns needed for model training while removing all personal identifiable information. The synthetic data generation system produces artificial transaction records that mimic the structure, distribution, and relationships of real data without containing actual customer information, thus enabling model training while protecting privacy.
Solution Approach 2:
The patent introduces a trusted intermediary entity that receives real transaction data, generates synthetic data, and distributes it to participating institutions. This intermediary acts as a mediator that enables collaborative model training across multiple institutions without any single institution having direct access to other institutions' sensitive data, thus preserving privacy while enabling collective learning.
2Object-affected harmful factors
If synthetic data is generated and used for model training, then privacy protection is improved, but data quality and model performance may deteriorate
Solution Approach 1:
The patent employs generative adversarial networks and other advanced synthesis techniques that carefully adjust and preserve critical data parameters such as statistical distributions, feature relationships, and pattern structures while removing PII. By maintaining these essential parameters, the synthetic data retains sufficient quality for effective model training despite being artificial.
Solution Approach 2:
The patent combines synthetic data from multiple participating institutions through federated learning, merging diverse datasets that all lack PII but collectively preserve various data patterns and characteristics. This aggregation of synthetic data from multiple sources compensates for potential quality limitations of individual synthetic datasets while maintaining privacy protection.
3Adaptability or versatility
If data is shared across multiple institutions for collaborative learning, then model versatility and robustness are improved, but data security and privacy risks increase
Solution Approach 1:
The patent segments the collaborative learning process into distributed components where each institution trains local models on its own data and shares only model updates or synthetic data rather than raw data. This segmentation allows collaborative learning across institutions while preventing direct access to sensitive data, thus maintaining security while achieving versatility.
Solution Approach 2:
The patent uses a trusted intermediary to manage data sharing and synthesis operations across institutions. The intermediary coordinates the creation and distribution of synthetic data, ensuring that no institution directly accesses another's sensitive data while enabling collaborative model development. This intermediary layer maintains security protocols while facilitating versatile model training.
Data Source
AI summary
Systems and methods for generating synthetic data using federated, collaborative, privacy preserving learning models are disclosed. In one embodiment, a method for generating synthetic data from real data for use in a federated learning network may include: (1) conducting, by a backend for a first institution of a plurality of institutions in a federated learning network, a transaction comprising transaction data; (2) generating, by the backend for the first institution, local synthetic data for the transaction data using a local synthetic data generating model; (3) sharing, by the backend for the first institution, the local synthetic data with a plurality of backends for other institutions; (4) receiving, by the backend for the first institution, global synthetic data from the plurality of backends for the other institutions; and (5) training, by the backend for the first institution, a local machine learning model with the global synthetic data.