Adjusting risk scores as consumers’ behaviors change has been made possible by the use of artificial intelligence (AI)-driven risk modeling that adapts to shifts in spending – and can help FIs spot and prevent fraud and criminal activity.
In fraud and financial crime prevention, deep learning has accelerated that development cycle. Automated Deep Behavioral Networks in particular represent a quantum leap forward within the field. This technology gives banks and financial institutions the tools they need to not simply save money but disrupt the entire fraud supply chain.
In terms of the efficiencies gathered, a comparison against the fraud detection rate of Featurespace's first-generation market-leading models revealed an uplift of 38 percent.
This is the nature of working with rapidly evolving technologies. Featurespace's mission is to fight fraud and financial crime on behalf of our customers. Morally, we feel an obligation to keep pushing our models further and further in pursuit of that mission. Tomorrow’s good work will obviate today’s good efforts.
Deep learning powered fraud prevention is now very real, and the following is a brief introduction to deep learning technology, how it is used to prevent fraud and why banking executives should be excited about its current state of development.
We have hit the limits of machine learning
Most banks’ fraud prevention tools today are built with machine learning (ML). That’s understandable because machine learning has done an excellent job at improving banks’ responses to fraud attacks.
ML-powered software can work in real time to assess the fraud risk of a given transaction. Just a few short years ago, that capability was good enough for most banks. Not today, however. Globally fraud is innovating at such speed and scale right now that traditional machine learning cannot keep up.
A traditional machine learning model works by applying logic to raw data to find a relevant signal in all that noise - this process is called feature engineering. The key word here is 'logic.' Whose logic does it apply? Our own.
Humans manually encode the things they think will be relevant for decision making into the machine learning model. And as we know all too well, human logic is rife with biases and blind spots. As soon as the model hits a problem its engineers had never considered, the model is out of its depth. The model is limited by the insight of its human creator.
That’s what holds machine learning back.
So, in order to outpace fraudsters – who thrive in the financial system’s corners and loopholes and places no one ever considered – banks need models that can learn much, much more deeply than humans can.
Deep learning pushes past those limits
Deep learning solves that learning problem by giving models the ability to extract relevant context from raw data independently of the logical insights of their creator. For transactions, that means empowering models to extract signals from the past as well as the present – i.e.: creating models with memory.
Language processing is a helpful analogy here.
Imagine I gave you a single word – e.g., 'knife' – and asked you to determine whether that word was threatening. Would you connect “knife” to assault, or would you connect it with spreading butter on your morning toast? Your interpretation will affect your assessment of whether the word is threatening. That’s a bit like what machine learning models experience.
Now, imagine you had a whole sentence or a whole paragraph to work with. That’s what deep learning does for the model. With the fuller context of a whole paragraph, you would be able to assess whether 'knife' is a threatening word with much more confidence.
In fraud prevention, 'knife' represents a single customer’s transaction, and the paragraph is that person’s history of financial transactions. If the transaction in question looks out of place or anomalous in that context, deep learning-powered fraud software can flag the transaction for the bank’s fraud team.
This is a capability that many banks are lacking today. As a result, they are forced to play catchup to fraudsters and organized criminals whose methods grow more and more sophisticated by the day.
Our goal now: Push the boundaries of deep learning
So, in the case of the customer mentioned above, why did Automated Deep Behavioral Networks outperform the older model? Because our field has so many smart, dedicated people who are pushing the limits of deep learning technology.
For those of us working in fraud and financial crime prevention, we understand that there is never a 'good enough' model. We must always innovate and always advocate for more R&D because that’s exactly what financial criminals are doing.
Here at Featurespace, our engineers and data scientists have developed new types of deep learning algorithm specifically for fraud prevention. For example, the memory cells in our Deep Behavioral Networks have been built to learn from, remember and understand human behaviors. This allows the network to recognize both individual, anomalous behaviors and sudden shifts in behaviors among whole groups of customers.
As our team experiments, learns, grows and challenges one another, our models will become more refined, more precise and more powerful.
There is a real sense of pride that comes from building something then building something better – and that is precisely our mission.
About Dr. David Sutton
Dr. David Sutton directs research and development (R&D) at Featurespace, driving the ongoing development of our world-leading machine learning and analytical technology. David joined Featurespace in 2015, having previously worked as a researcher at the Institute of Astronomy in Cambridge. He completed his DPhil in Astrophysics from the University of Oxford in 2010.
About Featurespace™
Every day we send out a free e-mail with the most important headlines of the last 24 hours.
Subscribe now