How Amazon Fraud Detector works

Amazon Fraud Detector builds a machine learning model that is customized to detect potential fraudulent online activities in your business. To get started, you provide your business use case. Depending on your business use case, Amazon Fraud Detector recommends a model type it will use to create a fraud detection model for you. In addition, it also provides insights into the data elements you need to provide as part of your business’s historical data. Amazon Fraud Detector uses the historical dataset to automatically create and train a customized model for you.

The automated model training process involves choosing a machine learning algorithm that detects fraud for your specific business use case, validating the data you provided, and performing data manipulations to improve model performance. After training the model, Amazon Fraud Detector generates model scores and other model performance metrics. You can use the score and the performance metrics to evaluate model performance. If needed, you can add or remove data elements from the dataset you provided for training and retrain the model to improve the model score.

After the model is created, trained, and activated, you need to configure decision logic, also known as rules, that tells the model how to interpret the data generated by your business, and assign outcomes for how to deal with interpretation of each activity. The outcomes can represent actions such as, approve or review the activity, or it can represent risk levels of the activity such as high risk, medium risk, and low risk.

A detector is a container that holds your model and the associated rules. You will need to create, test, and deploy the detector to your production environment.

The detector deployed in your production environment provides the fraud detection capability to your business applications. To perform fraud evaluation, the model compares all incoming data from your business activity with your business's historical data and uses its’ sophisticated machine learning algorithms with the rules you created to analyze the results and assign outcomes. With Amazon Fraud Detector, you can either evaluate data from a single business activity in real-time or evaluate data from multiple business activities offline.

Let us say you have a business that has online funds transfer as one of its activities. You want to use Amazon Fraud Detector to detect fraudulent requests for funds transfer, in real time. To get started, you will need to first provide Amazon Fraud Detector with data from past fund transfer requests. Amazon Fraud Detector uses this data to create and train a model that is customized to detect fraudulent requests for fund transfers. And then, you create a detector by adding the model and by configuring rules for your model to interpret the data. An example of a rule for online funds transfer activity can be, if the request for funds transfer is coming from xyz@example.com email address, send the request for review. In your business’s production environment, when a request for fund transfer comes in, the model analyzes the data that came with the request and uses the rule to assign the outcome. You can then take an action on the request depending on the assigned outcome.

Amazon Fraud Detector uses components such as, training dataset, model, detector, rules, and outcomes to provide your business with a fraud evaluation logic.

For information about the workflow you'll use for detecting fraud using Amazon Fraud Detector, see Detecting fraud with Amazon Fraud Detector