Charles Poladian

Jun 14th 2017

Machine Learning at the Cutting Edge of Fraud Prevention


Preventing fraud in a world of data breaches requires more than a handful of observant humans verifying each transaction. Machine learning is a powerful tool for fraud prevention and detection. Considering fraud’s impact on finance, e-commerce and insurance, advanced algorithms capable of scaling large pools of data are an increasing necessity for industries and consumers looking to thwart scammers.

The Power of Machines Who Can Learn

As the name implies, machine learning is a way for computers to learn without the need to regularly input new programs or updates. It’s a type of artificial intelligence using algorithms to sift through massive amounts of data. We take for granted email spam filters, but that’s a type of AI that prevents bad emails, from unverified accounts or untrustworthy actors, from ever reaching your inbox. These filters look for links that may lead to unsafe websites, email addresses that differ from known accounts of a company and language from the message itself to determine if it’s spam.

Similar algorithms can also be used as part of conversational AI systems. Northrop Grumman’s Synthetic Adaptable Intelligent Entity, or SAdIE, processes data on different topics to provide answers to any inquiry naturally.

Applying that same principle, algorithms are developed to detect potential fraud. You may have had a purchase flagged because it was for an amount, or from a location, that does not match your pattern of spending or history. That’s machine learning applied on a personal level and represents what’s happening across various industries on a much larger scale.

Preventing Fraud

Credit fraud alone cost consumers $16 billion in 2016. That impact is even greater for industries who lose on goods lost or sold, associated fees and loss of consumer trust. The Target credit-card breach of 2013, for example, affected 40 million customers. Target estimates the breach cost the company $252 million, although that number drops to $105 million after insurance reimbursements and tax deductions. After the breach, Target also paid $19 million to those affected by the hack.

As hacking and fraud becomes more sophisticated, so too must the systems in place to detect and prevent such attacks. Financial institutions, instead of just focusing on an individual’s history and behavior, can use a vast history of transactions to flag something as fraudulent. That same pattern recognition can be used for insurance claims.

Supervised machine learning requires training using different models. A linear-regression classification approach includes a set of data used to reach a “true” or “false” decision. Banks could flag any purchase over a certain value, the number of times a card was used outside of its country of origin or any other value to determine the probability of a fraudulent transaction. Companies have to train this model by using previous examples of fraud so the computer can learn how to detect these types of transactions in the future.

While a linear approach will work for the majority of fraud detection, neural networks can be applied when there are billions of transactions across hundreds of countries and millions of customers. Neural networks, or deep learning, closely mimic how the human brain makes decisions.

PayPal uses multiple machine-learning methods, including neural networks, to weigh thousands of features and create models along with potential outcomes to determine the trustworthiness of a customer. All of this is happening within seconds, which helps create a seamless transaction for a consumer while generating enough obstacles to flag a would-be fraudster.

As data breaches become seemingly more commonplace, AI can help even the playing field in the ever-evolving world of fraud.