AI vs. AI: How We Use Fake Data to Catch Real Fraud

In today’s digital world, financial fraud is growing more sophisticated, and so should our tools to prevent it. At the heart of modern defense systems are Neural Networks, a type of artificial intelligence (AI) modeled after the human brain. These complex systems have become a critical line of defense, helping to detect suspicious transactions in real-time. But like any other tool, its effectiveness depends on the data it learns from.

One of the most exciting and powerful types of neural networks is the Generative Adversarial Network (GAN). As you’ll see, GANs are starting to make a real difference in solving one of the biggest challenges in fraud detection: data imbalance.

The Problem of Data Imbalance

data-imbalance-problem

Imagine training a security dog with photos of 1,000 friendly faces and just 5 criminals. The dog probably wouldn’t learn much about what a criminal actually looks like. That’s essentially what happens when we train a machine learning model on highly imbalanced data.

In the financial sector, fraudulent transactions are rare — sometimes less than 1 in 1,000. This imbalance confuses our models. Since almost every transaction is legitimate, the model can learn to simply predict “no fraud” every time and still appear highly accurate on paper. But when it misses that one rare but critical fraudulent transaction, the cost can be huge for both consumers and businesses.

Solving the Problem: Old Methods vs. New AI

To solve this, we need to create more examples of fraudulent transactions for our models to study. For years, a common method for this has been SMOTE (Synthetic Minority Over-sampling Technique).

What is SMOTE? In simple terms, SMOTE looks at the few fraud examples it has, picks one, finds its nearest neighbors (other fraud examples), and creates a new, synthetic data point somewhere on the straight line between them. While useful, this method has its drawbacks. It tends to create simplistic, unrealistic data because it doesn’t truly understand the complex relationships between different data features (like time of day, transaction amount, and location). This can lead to noisy or redundant data, making it harder for models to perform well in the real world.

This is where GANs come in. Instead of relying on older techniques, we explored how GANs could help us generate more realistic and diverse synthetic fraud data to better train our models.

What are GANs?

gan-image

A GAN is like a game between two competing neural networks: a Generator and a Discriminator.

• The Generator tries to create fake data that looks completely real.
• The Discriminator tries to spot which data is fake and which is real.

First, the Discriminator learns the difference between real and fake data while the Generator’s learning is put on hold. Once the Discriminator’s learning has peaked, the Generator is allowed to learn how to make better fake data while the Discriminator’s learning is put on hold. They compete in a continuous back-and-forth process. Over time, the Generator gets so good at creating realistic-looking data that the Discriminator can no longer tell the difference. While GANs are famous for creating realistic human faces or art, they’re increasingly being used to solve complex problems with structured data, like financial records.

Using CTGAN to Generate Synthetic Fraud Samples

original-vs-generating-samples

There are many types of GANs, but when it comes to the tabular data found in spreadsheets (with columns for numbers, categories, and timestamps), most struggle. This is where CTGAN (Conditional Tabular GAN) is particularly useful.

CTGAN is specially designed for this kind of structured data. It understands how different features (like transaction amount, payment method, and user history) relate to each other, and it can generate highly realistic synthetic examples of rare cases, like fraudulent transactions.

By generating additional fake-but-realistic fraudulent data, we were able to:

• Balance our training data without simply duplicating rare fraud cases.
• Expose our model to more diverse and believable fraud patterns.
• Improve our model’s ability to detect fraud in the wild.

Instead of blindly guessing or overlooking fraud, the model became better at recognizing complex fraud signals that it never had the chance to learn from before.

Our Results

We used the CTGAN to generate 10,000 new synthetic chargeback records. Our original training data had about 150,000 fraud samples, including 20,000 real chargebacks. Instead of adding random data, we carefully added only chargeback-related examples to strengthen our model’s focus where it mattered the most.

This small, yet focused boost in the training data helped our model perform noticeably better:

• It became more accurate in identifying chargebacks.
• It was able to classify more chargebacks as fraudulent.
• The overall rate of chargebacks that the model missed capturing went down.

Even a 1% improvement in catching fraud can mean saving thousands of dollars in losses. This experiment demonstrated that adding the right kind of synthetic data, even just a small amount, can make a real difference in the model performance.

The Future of GANs in Finance

The use of GANs in finance is just getting started.

• Beyond fraud detection, they can help simulate rare economic scenarios, generate synthetic user behavior for testing, or create realistic test datasets where privacy is a concern.
• GANs could also help with stress testing, anomaly detection, or even market modeling.

As financial systems grow more complex and data privacy becomes increasingly important, generative models like CTGAN may become essential tools — helping us prepare for what we haven’t seen yet and catching what we might otherwise miss.

Author — Purva Natoo. The opinions expressed are those of the authors, based on their experience at Blackhawk Network.