How to detect payment fraud using machine learning?

Nowadays standards of financial security are being refined, and so are fraudsters’ methods of stealing data and money. Years ago criminals had to counterfeit client IDs, and today a person’s account password can be all it takes to break the bank.

Customer loyalty and conversions are affected in both environments, the digital and the physical. Javelin's Strategy & Research study found out that it takes 40+ days to detect fraud for brick-and-mortar financial institutions. Banks that provide online payment services are severely affected by financial fraud, this drives away more than 20 percent of individual customers.

So, the challenge for industry players is to make financial transactions more secure and improve the accuracy of fraud detection. Here comes machine learning which can be used for creating a fraud detection algorithm that helps in solving these real-world problems.

Types of internet fraud

Fraudsters have become very skilled in finding escapes so that they can steal more. Here are the most popular internet fraud types that we should watch out for:
- Email phishing (when hackers send fraudulent emails designed to trick people into falling for a scam);
- Payment fraud (attackers use stolen credit or debit card numbers to make an unauthorized transaction);
- Identity theft (faking users’ personal data to perform fraudulent activities).

Long before artificial intelligence widespread, security assistants were using a rule-based approach for fraud detection. But since machine learning appeared at all industry levels, companies have moved from rule-based fraud detection to ML-based solutions. In short terms, machine learning is the science of designing and applying algorithms that are able to learn things from past cases. It uses complex algorithms that iterate over large data sets and analyze the patterns in data. What is the difference between the old and new approaches?

The difference between ML and rule-based systems in fraud detection

The machine learning approach to fraud detection has received a lot of publicity in recent years and shifted industry interest from rule-based fraud detection systems to ML-based solutions. What are the differences between machine learning and rule-based approaches?

1. The rule-based approach. Fraudulent activities in finance can be detected by looking at on-surface and evident signals. Unusually, large transactions or the ones that happen in atypical locations obviously deserve additional verification. Purely rule-based systems entail using algorithms that perform several fraud detection scenarios, manually written by fraud analysts. Today, legacy systems apply about 300 different rules on average to approve a transaction. That’s why rule-based systems remain too straightforward. They require adding/adjusting scenarios manually and can hardly detect implicit correlations. On top of that, rule-based systems often use legacy software that can hardly process the real-time data streams that are critical for the digital space.
2. ML-based fraud detection. However, there are also subtle and hidden events in user behavior that may not be evident but still signal possible fraud. Machine learning allows for creating algorithms that process large datasets with many variables and help find these hidden correlations between user behavior and the likelihood of fraudulent actions. Another strength of machine learning systems compared to rule-based ones is faster data processing and less manual work. For example, smart algorithms fit well with behavior analytics for helping reduce the number of verification steps.

Even though the traditional process does not meet the contemporary requirements of security, its major disadvantage is the occurrence of false positives. This means completely normal customers just looking to make a purchase will go away from your business.

The judgment is dependent on individual training and transaction guidelines, which vary depending on the business. There will be high rates of false-positive if the employees reject every transaction above a certain risk threshold, or if it is cheaper to lose a sale than have a fraudulent transaction.
There are two most frequently used ML-based models for detecting fraudulent activities in financial transactions.

Fraud detection models

Supervised learning. These models are designed on tagged outputs. When a transaction occurs, it receives a ‘fraud’ or ‘non-fraud’ tag. Large amounts of tagged data are fed into the supervised learning model in order to train it in such a way that it gives a valid output. Also, the accuracy of the model’s output depends on how well-organized your data is.

Unsupervised learning. Such models are used to detect suspicious behavior in transactions which is new for the system. Unsupervised learning models include self-learning mechanisms that help detect hidden patterns in financial transactions. This model implies learning by itself, analyzes all incoming data, and tries to find the similarities and dissimilarities between transactions. This helps to detect illegal activities in a timely manner.

Both supervised and unsupervised models can be used separately or in synergy with each other to achieve maximum efficiency.

Why use ML in fraud detection?

Here are some factors for why ML techniques are so popular and widely used in industries for detecting frauds:

Speed. Machine learning is widely used because of its fast computation. It analyzes and processes data and extracts new patterns from it within no time. For human beings to evaluate the data, it will take a lot of time, and evaluation time will increase with the amount of data.

As we mentioned before, the traditional rule-based fraud prevention systems are based on written rules for permitting which types of actions are deemed safe and which ones must raise a flag of suspicion. It takes too much time to write these rules for different scenarios. And that’s exactly where ML-based fraud detection algorithms succeed in not only learning from these patterns it is capable of detecting new patterns automatically. And it does all of this in a fraction of the time that these rule-based systems could achieve.

Scalability. As more and more data is loaded into the ML-based system, the model becomes more accurate and effective in prediction. Unlike rule-based systems, a dedicated team of data science professionals must be involved in making sure ML-based algorithms are performing as intended, which implies constant modernization of the software product.

Efficiency. ML technology releases the burden of routine data analysis and repetitively searches for hidden patterns. Their efficiency is better in giving results in comparison with manual efforts. It avoids the occurrence of false positives which counts for its efficiency.

Due to their efficiency in detecting these patterns, the specialists in financial fraud detection could now focus on more advanced and complex patterns, leaving the low or moderate level problems to these ML-based algorithms.

Final thoughts

Since machine learning technology is a very popular instrument among financial companies as well as other industries’ experts, there is a big room for innovation. Experimentation with different algorithms and models can help any finance-related business in detecting fraud. ML-based techniques are undoubtedly more reliable than traditional human review and transaction rules.

To conclude, the machine learning solutions provide real-time processing of a large number of transactions, which ensures technology has a great future in the financial industry.


Author:
Egor Bulyhin, CTO at Smart IT, speaks on social media around the world focusing on software development strategy for Fintech, Healthtech, and other hi-tech areas. A former Project Manager and Lead Software Engineer, Egor practiced for more than 15 years on multiple IT projects before sharing his experience.