Detecting and Predicting Fraud with AI and ML

AI can detect and predict customer fraud, such as payment fraud and user account takeovers. Online retailers already detect many times of fraud on their websites and through their payment providers. The essential services and components of an e-commerce website can be monitored to detect fraud attempts. For example, the user registration and sign-in services can monitor for fake user account creation or credential theft. Brute force sign-in attempts can be tracked and flagged. Notifications can be sent to admins with appropriate levels of priority.

Policyholder fraud can be detected and predicted for domestic and commercial insurance. Predictive AI can use data provided at the point of quote to estimate the likelihood of fraud. Quotes can be declined or adjusted. Fraud risk information can also be provided to any subsequent claims investigation process for manual review.

Employee fraud can be catastrophic. Many small and medium -sized businesses are particularly exposed to internal fraud. AI can help detect theft and loss from employees early on. Payroll and expenses fraud such as timesheet manipulation, manipulation of employee databases and payment process abuse can be identified. Lower-level (but often significant) benefits fraud such as excessive or inadmissible sickness absence can also be highlighted by AI. Excess levels across individuals, teams or regions that then be suggested for further human review.

Machine learning can also help detect supplier fraud. Suppliers often have trusted access to internal business systems that can become compromised. Sometimes, supplier systems can themselves be compromised in attacks and used against you with supply chain fraud. This type of fraud can compromise payment systems, order and despatch processes and workforce management systems.

If generative AI is the lead singer that wows people by creating new content, machine learning is the bass guitarist in the background putting in the long hours of hard work to keep everything together.

Predictive fraud detection can feel like magic due to the complex and concealed nature of the algorithms.

So how does it all work?

Machine learning is, in essence, the process of using training data to determine how the various inputs affect the outputs. Any data source will contain a number of values (the attributes). Some of these attributes will have a significant influence on the output, while others will have little or none.

Traditional programming uses an imperative approach where the developer decides the steps of the algorithm ahead of time. While these algorithms can be highly complex, the programmer has effectively taught the solution to the computer.

Machine learning allows the computer to infer the correct algorithm for itself.

A portion of the starting data source can be used for testing the model. This is removed from the training set and used exclusively for validation. This allows confidence scores to be provided. Where the confidence in the result is low, human review can be suggested. Ongoing corrections for false positives or false negatives can continuously improve and refine the models.

Machine learning is particularly effective when the data is too complex or large for a human to reason about. If the solution to a problem cannot be succinctly expressed, machine learning may be able to outperform its human counterpart.

Fraud detection and prediction can be added to existing business processes. The fraud detection logic can be added as a discrete step in the pipeline. This approach allows the fraud detection results to be calculated in near real-time and directly affect the outcomes presented to users. For example by declining an order or sending immediate alerts.

An alternative approach is for the fraud detection routines to run alongside the existing pipeline in parallel. This approach has the advantage of not requiring any changes to existing logic and can be easier to implement. Data from the transactions can be processed by the fraud detection logic and flagged for review if required. This approach can work well when there is a natural delay in the business process, for example, between the receipt of an order and its dispatch.

Implementing AI fraud detection

The first step to adding fraud detection is to select the candidate data that might act as indicators. It is usually better to feed more data points into the training model than fewer. The model generation process should be able to work out which data points are good predictors of outcome. It is easier to refine a working model to use less data than to diagnose why models aren’t working well enough.

Many cloud data services provide hooks that can be used to trigger the AI functions. For example, Azure Event Grid can react to new incoming data and reliably trigger a processing function.

Training the machine learning models (and testing)

Training is the process of working out how much importance (or weight) the data points have in determining the outcome. Some data points might be very influential – some might even have a negative influence on the outcome.

An essential part of the training process is model validation. Typically, only a subset of the data available for training will be used for building the model (say, 80%). The remainder is then used for validation. This is useful because the training data has the expected results, so the testing subset can check for false positives and false negatives.

Both Azure and AWS have machine learning services available that are designed to accelerate the training, validation and implementation. Azure Machine Learning has a dedicated user interface (ML Studio) that can be used alongside an API for app developers. Amazon SageMaker is a similar environment for training, testing and deploying ML solutions. Both cloud platforms include MLOps tools to support deployment and manage changes to the model over time.

Continuous monitoring and verification of the models is critical to ongoing accuracy. Retraining models over time helps to ensure reliable and responsible predictions.