Independent AI review certifies best practice for fraud prevention solution

Machine learning (ML) is the most widely used method of artificial intelligence (AI). ML systems learn to recognize certain patterns or laws largely independently on the basis of examples. However, the cause-effect relationships in a trained model are not always open, i.e., the logic of how a result is derived from the initial information is often hardly apparent. This is why the metaphor of the black box is often used in public discussion. This is accompanied by the fact that ML systems are not only evaluated positively, but also critically, which forms an additional hurdle for a broader use of the technology.

Detailed analysis of the ML solution

As is generally the case with any use of models, responsible and compliant handling is particularly necessary with ML systems. As is so often the case, technical progress is also bringing regulators onto the scene: the EU Commission, for example, is currently working on a draft law to regulate the use of AI. To ensure that customers are on the safe side when using the fraud prevention solution AI:drian, the information service provider Experian, which also has a branch in Switzerland, commissioned Fraunhofer IPA to evaluate whether the ML component Transaction Miner was developed in compliance with all best practices and legal requirements and delivers comprehensible results. In addition to the EU draft, the review was based on a white paper from TÜV Austria and an AI test catalog from the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS, which point the way for future market standards. With the conclusion of a very detailed four-month review phase, this investigation has reached a positive result.

Certify legal compliance with independent AI audit

"We are seeing more and more ML systems being used in areas that the EU Commission has defined as critical, for example in self-driving cars or - as in this case - in checking consumers' intentions to commit fraud and creditworthiness," explains Professor Marco Huber, head of the Cyber Cognitive Intelligence (CCI) department at Fraunhofer IPA. "Not all details of the upcoming EU legislation are already known. However, what algorithms must look like that deliver fair, explainable and comprehensible results is not trivial, but it is also not witchcraft. Experian's Transaction Miner review gave us the opportunity to look under the hood of a product that is already in practical use and work on a standardized process that will be usable once the EU legislation becomes law."

Work made easier by more than 95 percent

The Fraunhofer IPA-tested Transaction Miner is the central component in the fraud prevention solution AI:drian, which Experian recently launched in the DACH region. With the help of the ML algorithms used, AI:drian is able to detect online fraud attempts more reliably than previous systems. For a pilot customer in the e-commerce industry, 99.9 percent of the transactions that the transaction miner identified as legitimate based on its training were not fraudulent and were accordingly not rejected. As a result, instead of around 7,000 order requests per month, only 300 had to be manually checked - a workload reduction of more than 95 percent. "New developments such as machine learning are being critically observed by the public," comments Martina Neumayr, Senior Vice President Credit Risk & Fraud Services at Experian DACH. "There are also examples of ML systems delivering questionable results. With Transaction Miner, we have developed a particularly powerful ML component that can also be used where the EU Commission sees a strong need for regulation, for example in the financial sector. This is precisely why we were particularly keen to act transparently and have the technology independently verified. We believe that the use of new technologies can only meet with acceptance if there are independent checks such as those carried out by Fraunhofer IPA. All the better if we have paved the way for a future standardized procedure through this pioneering achievement."

A wide range of offers around "Reliable AI

Audits like the one described are one of the core offerings of the CCI department at Fraunhofer IPA. One of the research focuses there is the topic of "Reliable AI". In this context, the IPA experts develop, among other things, methods for explaining and safeguarding AI systems. This also includes conducting audits and validating or qualifying AI systems as independent and legally compliant. The service offerings are industry-independent and are suitable for companies ranging from start-ups to corporations. 

Source: Fraunhofer IPA. More information about the product mentioned: www.experian.ch