How to Fight Fraud with Machine Learning?

Marcell Gogan  | No copyright attribution required).

When you are renting an apartment via Airbnb, hailing an Uber or paying for a cup of coffee in your favorite shop, you are sharing your personal data with a business.

For you, as a customer, the process is simple, secure and trustworthy.

What you may not think about it how these exchanges are handled by the business. Peer-to-peer marketplace platforms, financial institutions, insurance companies and healthcare providers constantly engaged in the ‘balancing act' of great customer experience vs. security.

All of them want to satisfy the consumer demand for fast and seamless payment experience. The trade-off, however, of such high trust levels is the possibility of fraud.

Machine learning fraud prevention solutions have entered the scene to address the growing gap between security and usability. Unlike traditional, largely, manual risk assessment systems, ML-powered systems are capable to detect sophisticated fraud in a matter of seconds without placing additional security burden on the users.

Fraud in Numbers (No copyright attribution required).

Security algorithms have become more sophisticated in the past few years, and so did hackers. According to Juniper Research, online credit card fraud is predicted to reach $25.6 billion by 2020. For comparison, in 2015 only $10.7 billion was lost to fraudsters.

The three major sectors under attack are:

  • eRetail – 65% of fraud valued at $16.6 billion.
  • Banking – 27% valued at $6.9 billion.
  • Airfare – 6% valued at $1.5 billion.

Payment fraud isn’t the only hazard for online businesses though. Over 62% of users have been affected by phishing and social engineering attacks in 2016, staged to get hold of their personal data.

The insurance industry has been another ‘darling’ among fraudsters with an estimated $80 billion lost to fraud in 2017 alone. Earlier, in 2015, around $60 billion Medicare coverage payments were identified as fraudulent or overpaid.

Lastly, we have the sharing economy, built entirely on the concept of trust. However, 67% of users decided against using some P2P services for the lack of trust in them. Most reported that they didn’t feel confident in the platform’s mechanism to protect the “shared” items; their personal data and safeguard them against fraudulent users.

Using machine learning for fraud detection can help mitigate such concerns and positively impact the business bottom lines as illustrated in this infographic:

AI in the Sharing EconomyAI in the Sharing EconomyAI in the Sharing EconomyAI in the Sharing EconomyAI in the Sharing Economy
Presented by Credit:

But What’s Wrong With The Current Fraud Detection Mechanisms?

Fraud detection remains a hand-operated process. Up to 83% of businesses conduct manual transaction reviews. Additionally, 90% of companies use a rule-based approach to identify suspicious transactions. Again, this process is done by human supervisors, meaning that it’s time-consuming and often ineffective.

The rule-based approach no longer works for three major reasons:

  • Rules are rigid. They do not adapt fast enough to predict and respond to new fraud types. Your business remains vulnerable to emerging threats. Rigid restrictions (e.g. limiting your business activities only to certain geographies) also reduce your profit and hampers customer experience.
  • You block “good customers”. The rule-based approach uses binary labeling of the transactions, meaning that they can be market as authentic or fraudulent. The problem, however, is that a lot of transactions often get labelled as false positives – a normal purchase is marked as suspicious and gets rejected based on the employee's judgment. As a result, your business loses a sale and undermines the lifetime value of your customers.
  • Rules require frequent updates, and thus resources. Scaling your rule set is a cumbersome and expensive process. They can quickly become a bottleneck for business growth.

Using machine learning in the payment industry eliminates the need for a manual overview, reduces the time spent on fraud detection and ultimately helps improve your entire setup.

Image Credit: PwC

Prevent and Identify Risks in Real-Time

Payment fraud usually follows recognizable patterns. It may take days for a human eye to spot the abnormalities within large datasets. Yet payment fraud machine learning solutions can spot unusual cues in a matter of minutes and take respective action immediately. Your systems can be further trained to handle fraud prediction and alert human staff whenever any suspicious activity is about to take place.

In this case, machine learning not only protects your business against fraud but also enables significant operational savings as fraud management procedures become streamlined and automated.

For example, Entropay – a virtual credit card provider, reduced the time spent on fighting fraud by 90% and increased their user conversion stats by 15% after implementing a machine-based fraud detection solution.

If you want to implement a real-time fraud prevention mechanism, it’s important to choose the right technology stack. The two popular frameworks for operationalizing Big Data are Apache Spark and Nginx. The latter one can handle four times more requests per second and requires less memory to operate smoothly, meaning that your systems will function more efficiently. Nginx is a better solution for high load websites, analyzing large amounts of data on a daily basis.

Take Advantage of Self-Improving Detection Mechanisms

Machine learning systems can be trained to improve their analysis over time. As the system "crunches" more data, it becomes capable to recognize more sophisticated patterns faster and make more educated predictions of certain outcomes.

It’s a known fact that most hackers tend to reuse parts of their code when creating a new threat. If your algorithm already knows how to recognize the original code, it could be easily trained to identify the ‘upgraded’ threat. Hackers may be getting smarter, and so does your system.

Educated Decision-Making with ML Analysis

Machine learning models also boast higher efficiency. They are capable to analyze billions of data points, accumulated from various sources.

Insurance fraud detection using machine learning can help providers distinguish the fraudulent claims with higher efficiency and as a result, offer reduced premiums for the honest consumers.

For example, Octo Telematics has developed a new risk assessment integration for the auto insurance companies. They offered to equip vehicles with IoT black boxes to record data about the driver's behaviour, send it for further machine-based analysis and help the insurance provider determine the best rates for different types of consumers. At the same time, the model will analyse the accumulated data and automatically indicate fraudulent claims.

Ultimately, machine learning can help business build “algorithmic security” applicable to a wide range of processes – payments, account registration, claims management, lending and so on.

By sending machines to do the meticulous job of analyzing billions of data points, organizations drastically reduce the chance of false-positive results; enable progressive fraud prevention improvement and become capable to predict and mitigate risks before those occur.

DISCLOSURE: The views and opinions expressed in this article are those of the authors, and do not represent the views of Readers should not consider statements made by the author as formal recommendations and should consult their financial advisor before making any investment decisions. To read our full disclosure, please go to:


Emerging Growth

Alliance Growers Corp

Alliance Growers Corp is a Canada based diversified cannabis company. The company is primarily focused on the development of Cannabis Botany Centres in Canada.