Fraud costs and attacks are on the rise in the financial services industry. Traditional methods to combat fraud either detect fraud after it happens or cannot keep pace with the volume and sophistication of the attacks. A better approach is to employ real-time fraud detection and prevention. The computational requirements of such an approach can be met with cloud-based, GPU-accelerated artificial intelligence.

Real-time action to prevent fraud is becoming increasingly necessary due to fraud’s growing impact on financial services institutions. A LexisNexis True Cost of Fraud Study: Financial Services & Lending released earlier this year found that the cost of fraud for U.S. financial services and lending firms has increased between 6.7% and 9.9% compared with before the pandemic. Every $1 of fraud loss now costs U.S. financial services firms $4.00, compared to $3.25 in 2019 and $3.64 in 2020.

One factor driving this growth is that more transactions are done via mobile apps. As the pandemic started, the FBI issued a warning that it expected that cyber actors would attempt to exploit mobile banking customers using various techniques, including app-based banking trojans and fake banking apps. The problem has gotten worse as the use of mobile banking increased significantly during the pandemic, and many people continue to use it now for its convenience.

Another conduit used to perpetrate fraud is to exploit different weaknesses throughout the customer journey. Everything from how users authenticate themselves to any of the numerous touchpoints a customer has with an institution can introduce exploitable vulnerabilities.

How AI can help

Financial services institutions are using artificial intelligence (AI) and machine learning (ML) to spot fraud in the making in real-time and prevent it from happening.

For example, most traditional fraud detection approaches use rules that flag suspicious transactions. The approach might look for online purchases from a suspicious location or a type of customer uncharacteristically spending above a certain level. Such rule-based platforms are inefficient because they rely on expected customer behavior, generating a large percentage of false-positive responses. An AI/ML approach can be trained on real customer behavior instead of relying on a set of rules.

Often such approaches can be complemented by focusing on the individual customer versus the collective expected behavior of similar customers. It learns every time a customer makes a purchase. It searches for activities and patterns to understand what that customer’s typical purchase behavior looks like to spot suspicious activity. It is also being applied to other areas of financial crime such as in the fight against money laundering. For example, AI-based Know Your Customer (KYC) frequently provides additional insights that improve visibility into potential risks associated with financial crimes.

Training ML models and running AI for fraud detection and prevention requires huge computational resources. Workloads can greatly benefit from elastic and scalable cloud-based, GPU-accelerated resources running optimized AI/ML algorithms, routines, and libraries. Marrying the right cloud and GPU technologies can provide the requisite scalability, faster and more efficient detection, and increased accuracy.

Teaming with the right technology partners

Assembling the various compute and software elements needed to do AI-based fraud detection at the scale of high-volume transactions major financial institutions experience every day is a complex task. Many organizations often do not have the time, money, or skills to undertake a modern fraud detection effort. They, therefore, need to find partners with the right technology and deep industry-specific AI expertise.

Microsoft and NVIDIA have been working together for years in the AI/ML arena. Their partnership in this arena goes back many years with the aim of infusing NVIDIA GPU technology on Azure to speed up entire AI/ML pipelines. And much of the technology, best practices, and methodologies they have jointly developed can be applied to fighting fraud.

This partnership has brought many innovations to market that make GPU acceleration available to more developers and businesses interested in using AI/ML. They offered Azure Machine Learning service as the first major cloud ML service to integrate RAPIDS, an open-source software library from NVIDIA. That allowed traditional machine learning users to easily accelerate their pipelines with NVIDIA GPUs. They also integrated the NVIDIA TensorRT acceleration library intoONNX Runtime. That enabled deep learning users to speed inferencing.

Work in these areas has continued. Last year, Azure announced support for NVIDIA’s T4 Tensor Core Graphics Processing Units (GPUs), which are optimized for the cost-effective deployment of machine learning inferencing or analytical workloads.

The true strength of the partnership is that the technologies are tightly integrated and optimized. For example, libraries are used to perform certain tasks to efficiently use GPUs. Installing and configuring these libraries takes time and effort. Azure takes care of pre-installing these libraries and setting up all the complex networking between compute nodes through integration with GPU pools. Additionally, by collaborating, NVIDIA and Azure have developed optimal configurations for GPU-accelerated AI workloads. That saves companies time and operational costs.

Tying this back to fraud detection, the cloud-based, GPU-accelerated artificial intelligence offered by Microsoft Azure and NVIDIA makes the needed computation resources available to institutes that want to adopt modern real-time fraud detection and prevention. And they get all the benefits of running AI in the cloud including the ability to scale up and out, security, fact interconnections, and more. In this way, such organizations can use the most sophisticated AI/ML-based approaches to better protect the institution’s and their customer’s assets.