Fight against Fraudsters
Covid’19 continues to have a growing impact on the global economy, leading to shut down of small businesses and layoffs of many employees. One thing that we might’ve ignored in this hassle is the spike in the number of scams over the last few months. Covid-19 scams have cost more than $13.4 million dollars according to the Federal Trade Commission. With an increasing impact on the economy and the convoluted nature of the fraudulent scams, it would be highly necessary to implement real-time fraud detection and prevention system.
In this note, we highlight the impact of the fraudulent scams on the economy, the key driving patterns that represent fraudulent behavior, a graphical proof of concept simplifying the nature of the problem and a blend of graph theory and risk scoring models as a solution to detect and prevent frauds.
Why should one read this?
The readers and banks especially can achieve a three point deliverable which when combined together can help them save millions of dollars.
- An in-depth consultation on graph manipulation tools that can scale to the needs of the company.
- Improved scalability and performance for operations associated with the increasing nature of transactional data
- A robust risk scoring model for detecting fraudsters and potential targets in the network
Impact of fraud
The risk associated with online banking transactions and payment firms has grown dramatically over a couple of years and the increasing scale, diversity, and complexity of fraud continue to hamper the organization’s ability to meet the evolving security needs.
As per the FTC report on AARP (a nonprofit organization in the United States), overall fraud losses were more than $1.9 billion last year, up from more than $1.4 billion in 2018, for a 28 percent jump. [Article for Reference]
Fraudulent behavior
After analyzing the impact, it would be in our best interest to understand the crux of the problem and how we could represent the complex intricacies of fraudulent activities. Here are some examples of the most common fraud actions faced by banks.
It is important to take into account the integrated behavior and transactional relationships between user accounts while detecting fraud. Graph modeling has proven to be one of the most prominent techniques in analyzing connected entities ( one of the major examples being social network analysis )
Intuitive way of Visualizing fraudulent transactions
Financial fraud is usually performed by organized parties or loosely connected criminals. The graphical component could come in handy in capturing these customer relationships and a variety of transactional behavior patterns. The graph approach can prove to be advantageous as it would be possible to track down the flow of money which is considered as the most essential component to uncover underlying fraudulent patterns in real-time. Lastly, a graph provides a highly intuitive visualization to understand the behaviors of clients.
An approach towards detecting and preventing frauds
The process highlights a three-folded objective:
1. Perform network analysis to identify case-specific business rules that detect frauds
It is very important to identify what type of fraud one wants to detect. The business rules would differ based upon the nature of the fraud.
Here, the analysis could be split into two segments: discrete analysis ( where we analyze user account behaviors) and connected analysis ( where we use graph techniques to study the relationships between the users)
We would like to divide the features for discrete analysis across three categories:
The recency of the transaction: Recording the time interval between two transactions
The frequency of the transaction: Number of transactions over a definite period of time
The amount associated with the transaction: Money associated with each transaction
Here connected analysis would mainly represent two steps: community detection and link prediction
Community Detection: We split a large graph into smaller subgraphs to identify a group of interacting nodes. It is used to understand the structure of large and complex networks
Link Prediction: We identify potential victims and prevent them from frauds. It is used to predict future possible edges in the network
2. Augment the analysis by building risk scoring models to prevent frauds
Probabilistic risk scoring algorithms and predictive machine learning models show great promise in determining the fraud risk of a node based on features like properties related to individual user accounts and their network connectivity.
Based upon the risk scores allotted to the links predicted by the link prediction algorithm, we can determine potential targets in the network. Any incoming transaction on a node with a high-risk score from outside the community is highly likely to be flagged.
3. Ensure dynamic scalable nature of the model
There are several graph wrangling tools and algorithms available today that can scale up to billions of nodes and edges i.e. it can handle the huge number of banking transactional data where, for instance, user accounts would be nodes and the transactions between them would be marked as edges.
Conclusion
This article acts as an effective starting point for an organization to deep dive into their business, understand and identify fraud specific business rules and implement the recommended solution with some tweaks if necessary to build a robust real time fraud detection & prevention system.
