Fraud Detection in the era of AI

Recent advances in computing and storage systems have empowered a new wave of Artificial Intelligence (AI) applications that take advantage of deep neural networks and deep learning (DL). In the past decade, the deep learning requires very large amounts of training datasets, which were hardly available and very expensive to manage. Nowadays, DL is used in a variety of applications, which expose exceptional performance and enable functionalities that were not possible few years before. As a prominent example, Google’s Alpha AI has recently defeated Korean grandmaster Lee Sedol in GO, who is considered a genius in this particular game.

One of the main characteristics of AI techniques is their ability to learn patterns in massive data sets, as a means of enabling systems that require minimum or even no human intervention. To this end, AI experts have to build quite complex analytics models, which are trained and evaluated using very large amounts of data. Based on these models, AI disrupts entire business domains by undertaking complex problem solving, increasing automation and eliminating human-mediated, error-prone processes. In these ways, it optimizes business processes, thereby ensuring that tasks are carried out in a safer and in a more reliable fashion.

Use of AI in Fraud Detection

The detection and alleviation of fraudulent transactions are one of the primary applications of AI in areas such as banking, insurance, and retail payments. Fraud detection and prevention is typically based on the automated discovery of high-risk fraud-related patterns across very large volumes of transactional datasets, including streaming transaction data that feature very high ingestion rates.

There are different types of patterns that are indicative of potential fraud. As a prominent example, human (e.g., customer) behavior patterns can lead an AI agent in altering a risk profile towards increasing the potential risk. In particular, transactions that occur in particular geographical areas might raise suspicion and increase risk. Likewise, transactions linked to high-risk parties (e.g., merchants without a good track record or a bad reputation) can be an indication of fraudulent activity. Other factors that can raise suspicion include the IP addresses of the parties involved in a transaction, relationships to suspicious persons or activity in social media, as well as the emergence of one or more transactions with unusual monetary value. An AI analytics model will typically consider all the different fraud-related indicators based on different weights and scores. To this end, the use of large amounts of training data enables the specification of parameters and weights for complex neural networks that can effectively score and classify fraudulent transactions.

One of the main trends in using AI for fraud detection is a shift towards predictive and preventive detection. The latter refers to the timely detection of fraud-related patterns i.e. before any fraudulent transaction occurs, as a preventive measure. Such a proactive detection requires systems that can automatically collect customer-related datasets (e.g., credit card transactions, mobile payment), while at the same time analyzing them nearly in real-time. Hence, real-time AI requires a real-time technical architecture, which can handle streams of high-velocity at a very low latency.

Another success factor for AI-based fraud detection is the presence of domain experts in the data scientist team. Such experts must have a strong knowledge of the fraud domain, which will help them build analytics models with appropriate parameters and weights for the fraud detection task at hand. Accordingly, they can improve the AI learning process through tweaking parameters and refining weights based on feedback from the models’ evaluation using real-life datasets. In this way, experts can develop fraud analytics algorithms with optimal performance. Moreover, they can use the evaluation results towards improving their knowledge about fraudulent attempts, through observing and detailing the characteristics of normal purchasing behaviors, while at the same time differentiating them from fraudulent processes.

Building Effective AI Systems for Fraud Detection

AI scientists that specialized in fraud detection employ the following best practices towards developing and deploying effective solutions:

• Combination of Different Types of Learning Models: Fraud detection is a complex issue, which makes it impossible to discover a single, “one-size-fits-all” set of analytics models and learning algorithms. In most cases, data scientists develop custom models, tailored to the problem at hand. In doing so, they have to combine multiple learning models and techniques, including supervised learning, unsupervised learning and reinforcement learning.
• Development of Dynamic Behavioral Profiles: In fraud detection systems customers and transactions should be clustered in behavioral profiles that are associated with different levels of risk. AI scientists need to create such profiles and accordingly specify rules that enable their dynamic updates as new transactions occur. The latter updates depend on dynamic information about the individuals, the merchant, the account and devices used in each transaction. Note that the creation of proper behavioral profiles requires knowledge about human behavioral analytics, which falls in the realm of social sciences.
• Train analytics models with huge amounts of data: Deep learning requires training based on very large amounts of data. In most cases, data availability is more important for the end result than the efficiency of the analytic model or algorithm used. One of the main reasons why giant IT enterprises like Google, Apple, and Facebook are able to build the most sophisticated AI systems is the fact that they leverage tremendous amounts of data for training and evaluation. Therefore, one of the challenges of AI in fraud detection is to collect and take advantage of large datasets. In this context, AI scientists and engineers are likely to take advantage of different sources of data that contain information about customers or merchants such as their public data available in websites and social media. Such data should be used over and above transactions’ data and customers’ profiling data available as part of processes like KYC (Know Your Customer). Nevertheless, during data collection, it’s always important to ensure compliance with ethical requirements and legal mandates, especially when using personal data.
• Domain models and domain knowledge: Domain knowledge is very important in all data science problems. AI systems for fraud detection are no exception to this role. The classification of a transaction or commercial interaction between merchants and customers as a fraudulent way depends on the context. A specific sequence of transactions that is suspicious in a given context, could be totally normal in a different setting. Likewise, the thresholds that signal anomalous behavior vary depending on the geographical context, the timing of the transaction and the profile of the customer. Overall, AI experts will have to build domain specific models rather than relying on general purpose models that may not be effective in given geographic, temporal and transactional contexts.

Like in many domains, AI in fraud detection promises to deliver exceptional automation and intelligence. This could allow the development of systems that are very effective in detecting fraudulent transactions, including predictive analytics systems that are able to identify fraud-related patterns and indicators even before actual fraud occurs. AI systems are currently used in conjunction with human intelligence in order to reduce manual tasks, save time and reduce costs. For example, an AI system could automatically classify 99% of potentially suspicions transactions and act upon them by asking for extra verification, leaving to humans the rest 1% of borderline cases. That’s certainly part of how AI is disrupting the finance, retail, and insurance sectors.