Fraud Detection Accelerator Using AWS SageMaker

Solutions Overview

Financial services organizations face growing risks from cybercriminals. High-profile hacks and fraudulent transactions undermine faith in the industry. As technology evolves, so do the techniques employed by these perpetrators, making the battle against fraud a perpetual challenge. Existing fraud detection systems often grapple with a critical limitation: relying on stale data. The newest tactics often can be seen in the data. That's where the power of operational data comes into play.

By harnessing real-time data, fraud detection models can be trained on the most accurate and relevant clues available. MongoDB Atlas, a highly scalable and flexible developer data platform, coupled with Amazon SageMaker Canvas, an advanced machine learning tool, presents a groundbreaking opportunity to revolutionize fraud detection. By harnessing operational data and leveraging the power of real-time insights, financial institutions can fortify their defenses against cybercriminals who seek to exploit vulnerabilities for illicit gains. MongoDB Atlas proves its strength as an operational data store, accommodating high-volume transactional data with exceptional performance and flexibility. Meanwhile, Amazon SageMaker Canvas empowers business analysts to leverage AI/ML solutions effortlessly, providing a no-code platform that brings the power of advanced analytics to their fingertips.

Reference Architectures

Below, you will find the architecture used to build this fraud solution. The architecture includes an end-to-end solution for detecting different types of fraud in the banking sector, including card fraud detection, identity theft detection, account takeover detection, money laundering detection, consumer fraud detection, insider fraud detection, and mobile banking fraud detection to name a few.

The architecture diagram illustrates model training and near real-time inference. The operational data stored in MongoDB Atlas is written to the Amazon S3 bucket using the Triggers feature in Atlas App Services. Thus stored, data is used to create and train the model in Amazon SageMaker Canvas. The SageMaker Canvas stores the metadata for the model in the S3 bucket and exposes the model endpoint for inference.

click to enlarge

Data Model Approach

The data is divided into two separate files: one containing identity information and the other containing transaction data. These files are connected through the TransactionID. It's important to note that not every transaction includes associated identity details.

Based on the above two datasets, we prepare a test join on the TransactionID, adding the target column as Fraud.

Data courtesy of Kaggle.

Source Table1: Transaction
TransactionID,
TransactionDT,
Card_no,
Card_type,
Email_domain,
ProductCD,
TransactionAmt,
Transaction_ID
Source Table2: Identity
TransactionID,
IpAddress,
PhoneNo,
DeviceID,
Location,
Name,
Address
Test Data:
TransactionID,
Card_no,
card_type,
Email_domain,
IpAddress,
PhoneNo,
DeviceID,
ProductCD,
TransactionAmt,
isFraud

Building the Solution

The detailed step-by-step guide to build this solution can be found in this Github repo. Below you will find an overview of those steps taken:

1. Set up the S3 bucket to which the MongoDB Atlas data needs to be exported.

2. Set up an MongoDB Atlas Cluster.

3. Set up Atlas App Services.

4. Set up the Amazon SageMaker domain.

Key Learnings

• Understand the use of Atlas Application Services and Atlas Charts to build products at scale.

• How MongoDB integrates natively with external services (such as AWS SageMaker, AWS S3) to provide even more powerful applications.

Technologies and Products Used

MongoDB developer data platform:

• Atlas Database

• Atlas App Services

• Atlas Charts

• Atlas Data Federation

Partner technologies:

• AWS S3

• AWS SageMaker Canvas

Authors

• Babu Srinivasan, Partner Solutions Architect at MongoDB

• Igor Alekseev, Partner Solutions Architect at AWS