Universal ML Model Pre-training for Fraud Detection Without Data Sharing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Financial institutions face challenges in generating effective Machine Learning models quickly for new products or features due to insufficient historical data, and existing solutions encounter data privacy issues by requiring sharing of data across banks.
Innovation Solution
A computerized method for generating a classification Machine Learning model in a cloud-based environment using isolated datasets from different environments, employing the extreme Gradient Boosting algorithm for continuous training across tenants without data sharing, and deploying the model in target systems with no training dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional ML model training is performed using local historical data, then model accuracy is improved, but deployment time is extended to 6-9 months
Solution Approach 1:
The system performs preliminary action by pre-training a universal ML model using aggregated historical data from multiple tenants before a new tenant needs the model. This pre-trained model serves as a foundation that can be quickly adapted to new tenants without requiring them to wait 6-9 months for their own historical data to mature, thus reducing deployment time while maintaining model accuracy through the use of pre-learned patterns.
Solution Approach 2:
The system introduces an intermediary approach by creating a universal model that acts as a mediator between the general knowledge from multiple tenants and the specific needs of individual new tenants. This intermediary model can be quickly fine-tuned for new tenants using their limited historical data, eliminating the need to wait for extensive local data accumulation while still achieving accurate results.
2Reliability
If data is shared across banks to train ML models, then model performance is improved, but data privacy issues arise
Solution Approach 1:
The system applies segmentation by dividing the data processing into separate phases: a universal model training phase that aggregates data from multiple tenants in a secure environment, and individual tenant fine-tuning phases that use only their own data. This segmentation allows the system to benefit from multi-tenant data diversity for improved model performance while maintaining data privacy by preventing direct access to individual tenant data by other tenants.
Solution Approach 2:
The system uses an intermediary approach by introducing a universal model as a mediator that consolidates learning from multiple tenants without requiring direct data sharing between tenants. The universal model acts as a secure intermediary that processes aggregated data while preserving tenant privacy, allowing each tenant to benefit from collective knowledge without exposing their sensitive data to other tenants.
3Adaptability or versatility
If a universal ML model is trained using data from multiple tenants, then model versatility is improved, but data privacy risks increase
Solution Approach 1:
The system segments the model training process into a universal model component that learns from aggregated multi-tenant data to provide versatility, and individual tenant-specific fine-tuning components that adapt to specific tenant requirements. This segmentation allows the universal model to achieve high versatility by incorporating patterns from multiple tenants while maintaining data privacy through the separate fine-tuning stage that uses only tenant-specific data.
Solution Approach 2:
The system applies local quality by allowing the universal model to have broad applicability across all tenants while enabling each tenant to customize the model with their own local data and requirements. This ensures that the model maintains high versatility at the universal level while adapting to the specific local needs and privacy requirements of each individual tenant through controlled fine-tuning.
Data Source
AI summary
A computerized-method for generating a classification Machine Learning (ML) model, in a cloud-based environment, is provided herein. The computerized-method includes building an ML model by using different isolated datasets from different environments: (i) identifying tenants of a service-provider by a base-activity; (ii) retrieving a set of features of objects from a database of each identified tenants to detect common features; (iii) using an object storage service in each tenant's environment to retrieve a dataset having the detected common features; (iv) training a ML model to classify objects on each retrieved dataset corresponding to a tenant from the tenants. The training of the ML model is a continuous training where the ML model continues training after each dataset, and then deploying a trained ML model in a target tenant system to classify objects. The target tenant system has no training dataset and no feasible training thereon.