How We Leverage Both Real-Time Detection and AI/ML Procedures for Best-in-Class Fraud Control
By now, almost anyone who deals with data as part of their day-to-day responsibilities is familiar with the terms Artificial Intelligence and Machine Learning. And in case you were on a leave of absence since the end of 2022 until now, the world has adopted an AI-supervised and reinforcement learning-based chatbot called ChatGPT, faster than any other online consumer application in history. There is a load of information online to help you learn more about AI and ML. At the end of the day, AI is the use of software to support complex activities like planning and problem-solving in ways similar to the human mind.
Machine Learning is a subfield of AI that applies mathematical models using significant computing resources to answer questions and identify insights within large data sets. There are many Learning Algorithms within ML. Three of the most popular are Supervised, Unsupervised, and Reinforcement learning algorithms. Algorithms are categorized based on their use of training or truth data as well as the types of their underlying computational algorithms.
So how do we leverage analytics including Artificial Intelligence and Machine Learning to optimize Fraud Management? Can AI/ML help achieve faster and more accurate Fraud Detection beyond real-time rules-based detection? These are questions we will discuss here.
Our journey at LATRO over the last decade providing RAFM solutions worldwide has led us to our current solution paradigm. In this paradigm, we reconcile the need to identify and block fraud as fast as possible while leveraging the advanced capabilities of AI/ML that provide insights and detection algorithms not possible with the cognitive limitations of even the smartest human analysts.
As Fraud Management professionals, our job is to control fraud loss. Telecom companies spend significant money building teams and implementing systems to prevent and mitigate losses. As we all know, perpetrators of fraud are smart, dynamic, innovative, and persistent professionals in their own right. Our challenge is to detect fraud attacks as quickly and accurately as possible in order to avoid or minimize losses. At the end of the day, our employers expect that their investment in our function will prevent losses much larger than the cost to detect and control them.
So, our goal is to detect fraud as fast and as accurately as possible. Can AI/ML technology help us to do this?
To examine this question, we are going to look at the case of International Voice Bypass or SIM Box Fraud. Essentially, it occurs when fraudsters exploit termination rates between telecom operators. This typically occurs in markets where the cost to terminate an international voice call is materially higher than the cost to terminate a local call. A common way to perpetrate this type of fraud is with a device called a GSM Voice Gateway or SIM Box. A SIM Box converts international voice calls carried over SIP connections to a local GSM, 3G, or 4G call on a mobile network using a local SIM Card.
Historically, Bypass Fraud Management systems based on test calls or CDR analysis were configured with a set of rules which examine data in near real-time and take advantage of behavior patterns that differentiate SIM Box activities from normal subscribers. These rules are often defined by Analysts who use a combination of their subject matter expertise and manual offline analytical tools to develop those rules. FM systems that use rules-based techniques are designed for processing large amounts of data and executing the rules as close to real-time as possible, thereby enabling fast and accurate detection.
However, there are limitations in this approach to achieving detection speeds fast enough to stop losses and prevent fraudsters from achieving their own profitability. The rule-based approach to detection has limitations imposed by our human cognitive capacities, and, by their nature, test calls and CDR analysis rely on transactions where fraud has already occurred. So detection results derived from even the fastest computational platforms will be “after the fact.”
At LATRO, we invented an approach where we apply near real-time computational resources to process transactions from the signaling layer of the network rather than the billing layer such as CDRs. In this approach, we can use the same rules and pattern-based technology to detect SIM Boxes faster.
So, how can we use AI/ML technology to take it to the next level?
We can use both Unsupervised and Supervised based learning models to find the bypass fraud that our limited rules-based approach misses. The success of applying AI/ML is very dependent on the capability of the software – that is the models and algorithms implemented, as well as the quality of the data. We often need to test different models and spend time cleaning our data to ensure the best results.
Using unsupervised learning, we can scan our data sets and find the outliers we didn’t see with our own expertise and manual tools. Then we can apply supervised learning models where we input training or truth data (that is previously confirmed fraud detections) to identify vectors or data correlations not apparent to even our cleverist analysts.
This all sounds great. But what’s the challenge?
The challenge is that this all takes time. Even with automation and massive computational resources, there is fraud occurring on the network while our AI procedures are executing and we are interpreting the results.
This means there is a point of tension. The point of tension is between real-time detection and AI/ML operating in non-real-time parallel procedures. In our experience today, we need to reconcile this tension and leverage both our rules-based real-time detection capability as well as our ever-evolving and improving AI capabilities.
At LATRO, our FM systems include both real-time, proactive rules and pattern-based detection as well as the capacity to apply AI and ML algorithms to support continuous assessment of the rules-based results. Our system users leverage both capabilities in order to constantly adapt to changing fraud behaviors and attacks. Our AI essentially helps Analysts become smarter in identifying and defining rules that can execute in real time and in some cases, block SIM cards even before fraudulent activities take place.
Let’s explore some examples. Our FM system called Versalytics contains a feature we call Suspect Behavior Analysis. This feature uses both unsupervised and supervised algorithms. Specifically, our unsupervised algorithm uses a mathematical model called DBSCAN which shows us density-based spatial clustering of outlier activities. The system can plot these on maps for Analysts to view. In addition, Analysts can feed the system confirmed past fraud detections or “truth data” to also execute supervised algorithms based on classification models. In this way, the Versalytics AI improves an Analyst’s ability to better define rules and pattern definitions.
What is even more interesting is that the signaling transaction data we use in our pattern-based patented Protocol Signature detection algorithm provides a far higher degree of dimensionality in the data available to the AI. This means there are exponentially more vectors available for the mathematical models, making them even more effective in learning as compared to simpler and smaller CDR or billable data sets.
AI is already revolutionizing RAFM performance within telecoms. Will it eventually replace the need for Analysts and rules-based detection approaches to Fraud Management?
At LATRO, we are not there today. Instead, we reconcile the strengths of both approaches in order to realize new benefits in our fight against fraud and achieve faster and more accurate fraud detection.
Where is AI being used in telecoms outside Fraud Management? There is a very large range of applications and use cases for AI/ML within telecoms. The telecom market is somewhat unique in that the amount of available transactional and reference data is huge. AI can leverage these data assets to drive value into the organization both for the telecom business itself as well as for customers. Prediction, Forecasting, Outlier Detection, and Segmentation are all common examples.
The use of AI in telecoms is crucial and it is only going to become more important as the threat of fraud grows. With the right AI-driven fraud management system in place, telecoms can better protect themselves and their customers.