Federated Learning Aggregation for Privacy-Preserving Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Payment product issuing organizations face challenges in continuously improving the accuracy of fraud prediction models due to the limitations of proprietary training data, which may violate privacy laws and operational impracticality when consolidating data across organizations.
Innovation Solution
A peer-to-peer network system for continuous model training, where nodes train and aggregate machine learning models using private datasets without revealing underlying data, leveraging a permissioned blockchain network and federated learning frameworks to securely share model weights and insights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If organizations consolidate their transaction data in one repository to improve model prediction performance, then model accuracy is improved, but privacy laws and organizational safeguards are violated
Solution Approach 1:
The system segments the centralized data repository into distributed data silos across multiple organizational nodes. Each organization retains its own transaction data locally rather than consolidating into a central repository, thus maintaining privacy compliance while enabling collaborative model training through the distributed network architecture.
Solution Approach 2:
The patent introduces an intermediary aggregation layer that coordinates model training across distributed nodes without direct data sharing. The aggregation server facilitates collaborative learning by collecting model parameters and gradients from participating organizations, enabling improved model accuracy while preventing direct access to sensitive transaction data.
2Reliability
If proprietary training methods are used to maintain organizational safeguards, then data security is maintained, but model prediction performance is limited
Solution Approach 1:
The system merges multiple proprietary training methods from different organizations into a unified federated learning framework. Each organization contributes its proprietary training approach to the collective model training process, allowing data security to be maintained through local training while model prediction accuracy improves through aggregation of diverse training methodologies.
3Measurement precision
If a centralized data repository is created to leverage diverse transaction data, then model training quality is improved, but operational impracticality increases
Solution Approach 1:
The federated learning system enables each organizational node to perform self-service model training using its own transaction data without requiring physical data movement or centralized storage infrastructure. Nodes independently train local model instances and contribute to the global model through parameter sharing, eliminating the operational complexity of managing a centralized data repository.
Data Source
AI summary
In some aspects, the techniques described herein relate to a method including: providing, on a peer-to-peer distributed network, an aggregation server; implementing, at a first node of the peer-to-peer distributed network, a machine learning (ML) model, wherein the ML model is trained on a first private data set; sending the ML model to the aggregation server; replicating the ML model to a second node of the peer-to-peer distributed network, wherein the ML model is retrained on a second private data set to generate a retrained model; sending the retrained model to the aggregation server; aggregating the ML model and the retrained model to generate an aggregated model; and replicating the aggregated model to the first node of the peer-to-peer distributed network, wherein the ML model is promoted to a production model at the first node of the peer-to-peer distributed network.