Improving financial services anomaly detection with Mphasis and Azure Quantum


Quantum-inspired optimization (QIO) improves and adds scale to traditional anomaly detection, reducing payment card risk in the financial services industry.

In 2019, fraudulent activity led to tens of billions of dollars in losses from payment cards alone. Payment card providers worldwide are constantly looking to improve their methods to detect and block fraudsters from exploiting their networks. Their mission is not only to protect the consumer but to also offer a trusted network to their business partners.

Anomaly detection is a key tool employed in this constant battle against fraud. While the exact methodologies differ based on industry specifics, having a state-of-the-art system in place is a goal for payment providers, as well as other companies with similar challenges. Mphasis, a global leader in building cutting-edge technology solutions for a variety of industries, is using Azure Quantum to innovate and develop the next generation of anomaly detection solutions, reducing the risk and impact of payment card fraud.

Developing improved, more scalable approaches to detect fraud

The process starts with deciding what data to include in an anomaly detection model. For example, a location change could indicate both legitimate travel or a stolen card. Additional data points like the date, time, and value of a transaction make the data rich enough to drive a clear decision on whether to block a transaction or not.

Finding the right balance between making card users feel secure and delivering an enjoyable shopping experience is crucial. Therefore, defining the right thresholds for anomaly detection requires careful consideration. It’s also important that any changes to anomaly detection software are made in an inclusive way that avoids certain groups being targeted or profiled. Fortunately, there are many more instances of legitimate transactions than fraudulent ones. However, this poses a technical challenge when creating a solution as this imbalance of data can skew a machine learning model. Through earlier innovation, Mphasis developed a solution that tackles this problem using quantum-related approaches.

Now, with Azure Quantum’s suite of advanced optimization solvers combined with scale of Azure, Mphasis has further expanded and improved their anomaly detection solution set. This is achieved through the added capabilities provided by Azure Quantum QIO solvers, which flexibly handle large optimization problems whilst automatically finding the best parameters to run on them. The result is a highly efficient solution ready to unlock impact at scale for a variety of complex anomaly detection problems.

Advantages accessible through Azure Quantum

The solution makes use of Azure Quantum to train an ensemble of Restricted Boltzmann Machines (RBMs), which are a type of artificial neural network. RBMs are highly effective at learning patterns from small datasets. During the learning process they can generate additional data to supplement the data that has already been provided. In addition, their built-in random variations make them particularly well suited to solving optimization problems.

The workflow below details the end-to-end approach used by the Mphasis team. Input data is converted to an equivalent binary form before it is fed into the RBMs for training.

During training, the goal of the RBM is to compute the weights of the neural network as accurately as possible. Accordingly, at each presentation of new data, the model should be able to classify more successfully whether the data is fraudulent or not. Training is an iterative process, which first requires updating the weights of the network and second doing probabilistic sampling to estimate values of the “hidden layer” nodes. These two steps are repeated until training is complete.

In earlier work, Mphasis had successfully expressed the probabilistic sampling step as an Ising model. With the problem already in a compatible format, it became a simple process of applying the quantum-inspired methods in Azure Quantum.

Using a publicly-available dataset of anonymized credit card transactions, Mphasis benchmarked a range of parametrized and non-parametrized solvers available through Azure Quantum.

In their experiments, Mphasis trained an ensemble of RBMs using the solvers and achieved their best results with Quantum Monte Carlo, which reached an F1 score of 0.9998 with a training time of 24 minutes. Measured out of 1.0 overall, an F1 score helps to characterize the success of a model at learning from the training data, by evaluating the accuracy, precision, and recall of new test data.

From research to the market

With a state-of-the-art approach that combines sampling improvements, ease of experimentation, and the scalability of the Azure cloud, Mphasis has the confidence that their solutions will consistently deliver results for a wide variety of problems. And, in a world where we are increasingly migrating away from physical cash and to more digital payments, it is essential that we can trust the payment providers and networks that we use. Anomaly detection is at the heart of creating trust between consumers, businesses, and payment providers.

“Partnering with Microsoft to develop and benchmark an Azure Quantum-based anomaly detection solution blueprint allows us to have more tools available to provide the best solution possible for our clients and address a wider variety of needs,” says Dr. Jai Ganesh, Senior VP of Research and Innovation at Mphasis. “We are excited by the scalability benefits Azure Quantum can provide compared to other available techniques allowing us to approach larger, more complex problems. Our team looks forward to applying Azure Quantum’s optimization solvers to our client’s most challenging optimization problems in a variety of industries over time. As quantum computers mature and the offering in Azure Quantum grows, there is an almost unlimited set of future possibilities to further enhance this and other solutions.”

Get started today

Learn more about Mphasis’ Anomaly Detection solution using Azure Quantum.

Explore how to leverage powerful quantum-inspired optimization capabilities and get started on your own quantum journey with Azure Quantum.

See how others like Mphasis are participating in the vibrant researcher, quantum thought leader, academic, and solution partner community reflected in the Azure Quantum Network.

Nilson Report

The post Improving financial services anomaly detection with Mphasis and Azure Quantum appeared first on Microsoft Azure Quantum Blog.

Read More


Generated by Feedzy