Neural Network Fraud Detection System Using Synthetic Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting fraudulent transactions are inefficient due to reliance on manual review and predefined rules, which are burdensome, time-consuming, and inadequate for detecting complex or new patterns, and neural networks trained on limited labeled data struggle to classify unlabeled transactions accurately.
Innovation Solution
A system utilizing multiple neural networks, including a classifier, generator, and discriminator, to generate labels for unlabeled data, create fake data, and classify interactions as real or generated, thereby reducing manual effort and enhancing the ability to detect fraudulent transactions by training on a broader dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual review is used to identify fraudulent transactions, then detection accuracy can be maintained, but the process becomes burdensome, time-consuming, and expensive
Solution Approach 1:
The system performs preliminary automated classification of transactions using neural networks before manual review. By pre-processing and filtering transactions through multiple neural networks (classifier, generator, discriminator), the system prepares data in advance, allowing manual reviewers to focus only on ambiguous cases rather than reviewing all transactions from scratch.
Solution Approach 2:
The patent introduces neural networks as intermediary systems between raw transaction data and manual review. These networks act as mediators that initially assess transactions and only flag uncertain cases for human review, reducing the volume of manual work while maintaining detection accuracy through the collaborative human-AI system.
2Productivity
If only a small portion of transactions are manually reviewed and labeled, then manual effort is reduced, but sufficient training data for machine learning becomes unavailable
Solution Approach 1:
The system uses a generator neural network to create synthetic copies of fraudulent transactions. By generating artificial fraud cases that mimic real fraud patterns, the system expands the limited labeled training data into a larger synthetic dataset, enabling more comprehensive model training without requiring proportional increases in manual labeling effort.
Solution Approach 2:
The patent transforms the nature of training data by converting unlabeled transactions into labeled data through the neural network classification process. By changing the state of data from unlabeled to labeled through automated assessment, the system effectively increases the quantity of training data available without direct manual intervention for each case.
3Reliability
If predefined rules are used to detect fraudulent transactions, then known fraud patterns can be identified, but new patterns employed by fraudsters cannot be detected
Solution Approach 1:
The system replaces static predefined rules with dynamic neural networks that continuously learn and adapt. The neural networks are trained on transaction data and can evolve their detection capabilities over time, automatically adjusting to new fraud patterns as they emerge rather than requiring manual rule updates for each new threat.
Solution Approach 2:
The patent implements feedback loops where the neural networks are continuously trained on new transaction data and labeling results. This feedback mechanism allows the system to learn from both confirmed fraud cases and false positives, progressively improving its ability to detect emerging fraud patterns while maintaining detection of known patterns.
4Productivity
If classifiers are trained on limited labeled transactions, then training resources are conserved, but the ability to detect complex patterns is insufficient
Solution Approach 1:
The generator neural network creates synthetic copies of complex fraud patterns, expanding the limited training dataset. By generating additional realistic fraud scenarios that mirror complex patterns, the system enables the classifier to be trained on more diverse and comprehensive data without proportionally increasing manual labeling resources.
Solution Approach 2:
The system combines multiple data sources and types into a composite training dataset, including real labeled transactions, synthetic generated fraud cases, and potentially unlabeled data processed through the neural networks. This composite approach enriches the training data with diverse patterns while maintaining training efficiency.
Data Source
AI summary
A method for detecting fraudulent interactions may include receiving interaction data, including a first plurality of interactions with (first) fraud labels and a second plurality of interactions (without fraud labels). Second fraud label data for each of the second plurality of interactions may be generated with a first neural network (e.g., classifying whether each interaction is fraudulent or not). Generated interaction data and generated fraud label data may be generated with a second neural network. Discrimination data for each of the second plurality of interactions and generated interactions may be generated with a third neural network (e.g., classifying whether the respective interaction is real or not). Error data may be determined based on the discrimination data (e.g., whether the respective interaction is correctly classified). At least one of the neural networks may be trained based on the error data. A system and computer program product are also disclosed.