AI-Enabled Auditors: Finding Fraud Using Deep Autoencoder Networks

Accounting fraud has long eaten into the revenue of some businesses, but auditors are enlisting a new defensive tool: artificial intelligence.

A typical organization can lose 5 percent of its annual revenue to fraud, according to an estimate from the Association of Certified Fraud Examiners. Businesses are putting AI on the task of anomaly detection in an effort to staunch losses.

'Organizations are looking at what they can do in this space from a forensic point of view,' said Marco Schreyer, a researcher with the Deep Learning Competence Center for the German Research Center for Artificial Intelligence, at last month's GPU Technology Conference.

Schreyer's research focuses on anomaly detection, unsupervised deep learning and accounting fraud.

In addition to lost revenue, financial fraud can create a reputational black eye for companies, he said.

Companies continue to accelerate the digitization of their business processes, affecting their enterprise resource planning software and leaving behind an audit trail, Schreyer said.

ERP systems collect vast quantities of electronic journal data - the formal accounting entries made to record business transactions - on accounts at a granular level. But fraudsters generally deviate from ordinary system usage or accounting patterns.

'Now, using trained neural networks, we're looking at the flow of transactions and how they are captured in systems,' Schreyer said.

The Deep Learning Competence Center sought to flag accounting anomalies in data entries by training several deep auto-encoder networks using NVIDIA's DGX-1 system.

The model was then tested on two different real-world accounting datasets. The results showed the trained network was able to flag anomalous data entries for audit.

'The more you can contrast those networks, the better you can track the data,' Schreyer said. 'This approach can be used to detect anomalies in different spaces, not just finance.'