Presenters
Source
Fighting the $48 Billion “Friendly” Problem: How AI and Cloud-Native Tech Are Redefining Fraud Detection 🚀
In the world of financial services, not all fraud involves shadowy hackers in hoodies. Sometimes, the threat comes from a legitimate customer making a legitimate purchase. Gautham Paspala, a Stack Software Engineer at ServiceNow, is tackling one of the most expensive and elusive challenges in the banking industry: Friendly Fraud.
In a recent deep dive, Gautham shared how his team leverages Temporal Graph Attention Networks (GATs) and Transformers on a cloud-native architecture to stop fraudsters in their tracks.
💸 The $48 Billion Blind Spot: Why “Friendly” Fraud Isn’t Friendly
Friendly fraud occurs when a real customer makes a purchase and later disputes it with their bank to get a refund, despite receiving the goods. This isn’t just a minor nuisance; it is a massive financial drain.
- The Scale: Friendly fraud accounts for 60% to 70% of all disputes, costing the industry over $48 billion annually.
- The Challenge: Unlike traditional theft, the transaction itself looks perfect. The card is valid, the holder is verified, and the delivery is successful. The “fraud” only happens weeks or months later during the dispute process.
- The Failure of Legacy Systems:
- Rule-based systems (the “old school” way) only catch 23% to 31% of cases because they rely on rigid, predictable thresholds.
- Traditional Machine Learning performs slightly better (38% to 45%), but it struggles to understand the complex, evolving relationships between users and devices.
🧠 The Framework: A Two-Pronged AI Attack
To solve this, Gautham’s team developed a framework that combines two cutting-edge techniques to spot patterns that humans and simple algorithms miss.
1. Temporal Graph Attention Networks (GATs) 🕸️
This component models the entire payment ecosystem as a massive graph. It connects 50 to 100 million cardholder accounts and 15 to 25 million merchant entities.
- The Tech: The system tracks 10 to 50 billion edges (connections), with 100 to 500 million new connections added daily.
- The “Attention” Secret: Not all connections matter equally. The Graph Attention mechanism assigns weights to relationships. If multiple accounts share a single device or IP address, the model flags it as a coordinated fraud ring.
2. Transformer-Based Sequence Models 🤖
While GATs look at the “who and where,” Transformers look at the “when and how.”
- Behavioral Patterns: The model treats a cardholder’s history as a sequence, much like a sentence in a book.
- Automated Feature Engineering: It automatically identifies which past transactions suggest a user is systematically gaming the dispute process, removing the need for engineers to write manual rules.
🏗️ Scaling to the Stratosphere: Cloud-Native Infrastructure
Building a model is one thing; running it at the scale of a global bank is another. Gautham emphasizes three cloud-native pillars that make this possible:
- Distributed Processing: Using distributed graph databases with partitioned storage allows the team to parallelize work across billions of edges.
- Kubernetes-Based Autoscaling: This ensures the system handles massive traffic spikes without manual intervention, maintaining 99.9% uptime.
- Segment-Level Calibration: Instead of one-size-fits-all, the team tunes detection thresholds for specific merchant categories. This drastically reduces false positives (annoying a real customer) while minimizing financial loss. 🎯
📈 The Results: Better Detection, Less Friction
The impact of moving from legacy rules to this AI-driven approach is staggering:
- Detection Accuracy: The GAT and Transformer approach reaches 67% to 78% detection, a 40% to 65% relative improvement over traditional ML.
- Operational Efficiency: The system achieves a 35% to 45% reduction in false positives. This means fewer legitimate customers face card blocks and dispute teams have a much lighter workload.
- Metrics of Success: The team focuses on Precision, Recall, and AUC scores to ensure the model remains robust and independent of shifting thresholds.
🤝 The Human-in-the-Loop Advantage
Gautham is clear: automation and human expertise are not in competition.
The system uses a Real-time Human-in-the-Loop feedback cycle. When the model flags a case, fraud review teams make the final call. That decision is fed back into the training pipeline. This creates a self-improving loop where the model constantly adapts to new fraud tactics while increasingly protecting legitimate customers from unnecessary friction. 🛡️
🎯 Final Thoughts
By combining the relational power of Graph Attention Networks with the sequential intelligence of Transformers, and deploying them via Kubernetes and distributed databases, ServiceNow is proving that even the most complex “friendly” fraud can be managed.
The result? A safer financial ecosystem that is scalable, maintainable, and smarter every single day. 🌐✨