This post has been republished via RSS; it originally appeared at: Microsoft Tech Community - Latest Blogs - .
In a typical case of online fraud, the thief makes multiple transactions, leading to a loss of thousands of dollars. That's why fraud detection must happen in near real-time.
This article presents a solution that uses Azure technology to predict a fraudulent mobile bank transaction within two seconds. We've built it with customers.
Read the article here:
Let's dig into the architecture:
An event-driven pipeline ingests and processes log data, creates and maintains behavioral account profiles, incorporates a fraud classification model, and produces a predictive score. Most steps in this pipeline start with an Azure function. A model training workstream combines on-premises historical fraud data and ingested log data. Azure Data Factory orchestrates the processing steps. We use Azure Logic Apps to connect and synchronize to an on-premises system to create a fraud management case, suspend account access, and to generate a phone contact.
In the article you'll find:
- Information about the top challenges: Rare instances of fraud and rigid rules.
- Operational context: The key questions we asked and how fraud is committed in the operational environment.
- Compromise matrix: See the methods used, data taken, and patterns for several types of fraud, including Credential, Device, Financial, and Non-Transactional compromises.
- A detailed dataflow of the above architecture.
- Data pipeline and automation: What happens in the two seconds, in order to catch the compromise.
- Event processing: Architecture and dataflow that explains in detail the fundamental interactions for an Azure function within this infrastructure.
- Data pre-processing and JSON transformation.
- Near real-time data processing and featurization with SQL Database.
- Event schema management.
- Feature engineering for machine learning.
- AutoML: It automates the time-consuming, iterative tasks of machine learning model development.
- Data imbalance: In a fraud dataset, there are many more non-fraudulent transactions than fraudulent transactions.
- Model training with a code sample!
- Model evaluation: The account-level metrics are described in a table.
- Model operationalization and retraining.
- Components: Direct links to all the Azure services used in this solution.
- Technical considerations: Skill sets and Hybrid operational environment.
- Security considerations: Includes a Networking Security Architecture and a security baseline recommendations matrix.
- Scalability considerations.
You can find the article here, on the Azure Architecture Center:
Special thanks to the Engineers who wrote this:
- Kate Baroni
- Michael Hlobil
- Cedric Labuschagne
- Frank Garofalo
- Shep Sheppard
And thanks also to our editor/tech writer, Mick Alberts.
Remember to keep your head in the Cloud!
Ed