Graph Neural Network Modeling for Real-Time New-Account Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems struggle to effectively detect and prevent new account fraud (NAF) in online systems, particularly in content delivery networks (CDNs), as fraudsters obscure their identities using fake or stolen email addresses and IP addresses to create multiple accounts, making it difficult to identify fraudulent patterns.
Innovation Solution
A graph neural network (GNN) with an attention mechanism is used to analyze email and IP address data, mapping new account registrations into connected graphs to identify similarity cues and detect fraud cliques by leveraging the k-nearest neighbors of a query node vector, generating a probability of fraud through a self-attention output layer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional email-based fraud detection is used, then the system can identify some fraudulent accounts, but fraudsters can easily create fake email addresses to mask their identity and evade detection
Solution Approach 1:
The patent transitions from analyzing single-dimensional email data to multi-dimensional graph-structured data that includes emails, IP addresses, device identifiers, and temporal relationships. This dimensional expansion enables the system to detect fraud patterns that span multiple data dimensions, making it harder for fraudsters to evade detection by manipulating single data points.
Solution Approach 2:
The patent creates a composite detection model that combines multiple data types (emails, IPs, devices, timestamps) into a unified graph neural network framework. This composite approach integrates heterogeneous data sources to form a more robust detection system that can identify fraud patterns across different data modalities, reducing the effectiveness of single-point manipulation by fraudsters.
2Measurement precision
If the system analyzes detailed registration data to improve fraud detection, then detection accuracy improves, but processing time and computational complexity increase
Solution Approach 1:
The patent pre-computes and stores graph embeddings for registration data during the registration process itself, rather than performing heavy computational analysis only during fraud detection. This preliminary action prepares the data in advance, enabling fast real-time fraud detection by comparing pre-computed embeddings against fraud patterns, thus reducing processing latency while maintaining high detection accuracy.
Solution Approach 2:
The patent replaces traditional rule-based fraud detection mechanics with a graph neural network-based intelligent system. The GNN automatically learns fraud patterns from graph-structured data and performs detection through neural network inference, which is computationally more efficient than exhaustive rule evaluation, thereby reducing processing time while improving detection accuracy.
3Reliability
If the system uses complex graph neural network models to detect fraud patterns, then detection capability improves, but system complexity and implementation difficulty increase
Solution Approach 1:
The patent segments the fraud detection system into distinct modular components: graph construction module, embedding generation module, and fraud classification module. Each component handles a specific aspect of the detection process, making the overall complex system more manageable and easier to implement. This segmentation allows teams to develop and optimize each module independently while maintaining high detection capability.
4Reliability
If the system monitors all new account registrations in real-time, then fraud detection coverage improves, but computational resources and processing load increase
Solution Approach 1:
The patent applies fraud detection selectively based on risk indicators rather than uniformly to all registrations. The system identifies high-risk registrations using preliminary filters (such as multiple registrations from same IP, suspicious email patterns) and applies the computationally intensive graph neural network analysis only to these cases. This partial action approach maintains high detection coverage for fraudulent accounts while significantly reducing overall computational resource consumption.
Data Source
AI summary
A method executes upon receiving data associated with a registration. In response, an encoding is applied to the data to generate a vector. The vector indexes a database of such vectors that the system maintains (from prior registrations). The database potentially includes one or more node vector(s) that may have a given similarity to the encoded node vector. To determine whether there are such vectors present, a set of k-nearest neighbors to the encoded node vector are then obtained from the database. This set of k-nearest neighbors together with the encoded node vector comprise a virtual graph that is then fed as a graph input to a Graph Neural Network previously trained on a set of training data. The GNN generates a probability. If the probability exceeds a configurable threshold, the system outputs an indication that the registration is potentially fraudulent, and a mitigation action is taken.