ML-Driven New Account Fraud Early Detection System
Software Engineering Institute
The Australian Competition and Consumer Commission's (ACCC) through “Scamwatch” (scamwatch.gov.au) reported that in 2020, Australians lost a reported $851 million to fraud, up by 23.1 per cent compared to 2019. Investment frauds accounted for more than a third of total losses. The greatest challenge faced by online financial products is the increase in identity theft. As reported by the Australian Institute of Criminology (AIC), one in four Australians have been a victim of personal identifiable information misuse of some sorts. As rule-based detection systems are becoming less effective, the on-going challenge of fraud and financial crime requires innovative and effective approaches.
In response, we focused our R&D effort on detecting new account fraud activities that usually follow identify theft incidences or synthetic identity creation and use. The explicit business goals were to reduce false positives rate, improve true fraud detection rate, and enable the business to detect fraud cases before any monetary outflows by striking the fraudsters earlier in the Fraud “Kill-Chain”.
As part of our data transformation process, we leveraged the capabilities of graph databases to model the customer “360” view of existing relationships, and to implement “guilt-by-association” logic to unearth fraud networks no matter how deep or how hard they try to hide. Another key technical demand was the application of the imbalanced data handling techniques during model tuning process when facing 1000:1 normal to fraud data ratio.
When viewed from the business perspective, we were able to achieve over 80% reduction in false positive cases, thereby removing a significant portion of negative impact to customer relationships. The model also resulted in more true positive predictions. In some cases, the model was able to catch fraudsters’ attempts to evade existing detection systems trying to launder multi-million-dollar assets.
The sharing of successful adoption of some of the front-line non-proprietary techniques we applied in our new account fraud detection research will lift all boats and help our industry peers to collectively defend against organized criminal groups, opportunistic fraud attacks and other illicit actors; restricting their funding sources, diminishing their overall capabilities and ultimately deterring on-going fraud and financial crimes to the benefit of Vanguard and the wider financial services sector.
Will Li is a senior technical leader in risk and security space at Vanguard. His current focus is on promoting the adoption of analytics and machine learning across the many sub-domains of enterprise risk, security, and fraud management. Prior to that, Mr. Li has had a long and diverse career in planning, designing, and building infrastructure, cloud, and global networks spanning two decades. Outside of work, he also enjoyed adjunct teaching networking and security. Mr. Li holds two masters’ degrees in technology fields and is currently pursuing his PhD degree in Engineering. He has published numerous peer-reviewed articles in high-impact scientific journals in the field of data analytics and machine learning applied to cyber security and health sciences.
Jose Martins is currently the Senior Fraud Detection and Monitoring lead at Vanguard Australia. In his current role he is responsible for developing and maintaining fraud detection, investigation, and mitigation strategies to support the operation of several online trading platforms. Jose is an avid advocate of advanced analytics using descriptive, predictive, and social network techniques to uncover fraudulent and criminal activity. Mr. Martins brings an extensive background in complex investigations and signals intelligence collection and analysis from working a decade in law enforcement in Brazil. With a bachelor's degree in international commerce and finalizing a graduate degree in Fraud and Financial crime Jose is aiming for a formal degree in data science.
Part of a Collection
FloCon 2022 Assets
This content was created for a conference series or symposium and does not necessarily reflect the positions and views of the Software Engineering Institute.