Fraud Detection System Using Bi-Directional Recurrent Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The rapid growth of e-commerce has led to increased online fraud, with existing fraud detection systems struggling to accurately and timely identify fraudulent transactions, particularly those involving a 'burst' of fraudulent activities that may not be evident from initial transaction data alone.
Innovation Solution
A system that generates a second risk score using a bi-directional recurrent neural network, incorporating transaction data both before and after the initial transaction, to classify transactions as potentially fraudulent and trigger fraud mitigation actions, such as cancellation or verification requests, before fulfillment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a first risk score is generated based only on transaction data received prior to the transaction request, then the fraud detection can be performed quickly at the time of transaction, but the accuracy is insufficient to detect fraudulent activities that become evident only after the transaction is initiated
Solution Approach 1:
The system performs a preliminary fraud assessment using a first risk score based on historical transaction data before the transaction is initiated. This preliminary action enables quick initial filtering of potentially fraudulent transactions while preserving the ability to perform further analysis. The system prepares fraud detection infrastructure in advance and uses available data upfront to make initial risk assessments.
Solution Approach 2:
The system implements a feedback mechanism where transaction data received after the initial risk score is generated is fed back into the fraud detection process. A second risk score is generated incorporating this new information, allowing the system to refine its fraud assessment. This feedback loop enables the system to detect fraudulent patterns that emerge after transaction initiation while maintaining timely detection capability.
2Reliability
If traditional fraud detection methods are used that rely only on pre-transaction data, then the system complexity remains low, but the system cannot detect burst fraudulent activities that manifest after transaction initiation
Solution Approach 1:
The fraud detection process is segmented into distinct phases: initial risk assessment using pre-transaction data, and subsequent risk assessment using post-transaction data. Each phase uses appropriate data sources and analysis methods for that specific time period. This segmentation allows the system to handle different types of fraud indicators that appear at different stages without requiring all complexity simultaneously.
Solution Approach 2:
The system dynamically adapts its analysis approach based on the transaction stage. Initially, it uses simpler pre-transaction data for quick assessment. As the transaction progresses and additional data becomes available, the system dynamically incorporates this new information to generate updated risk scores. This dynamic approach enables detection of evolving fraud patterns while managing system complexity through staged analysis.
3Object-affected harmful factors
If all transactions are automatically canceled when fraud is suspected, then fraudulent transactions are prevented, but non-fraudulent transactions are incorrectly canceled increasing false positives
Solution Approach 1:
Instead of automatically canceling all transactions with elevated risk scores, the system applies partial action by implementing verification requests for transactions with moderate risk scores. Only transactions with high confidence fraud indicators result in automatic cancellation. This graduated response reduces false positives while still preventing confirmed fraudulent transactions, allowing legitimate transactions with ambiguous risk factors to complete after verification.
Data Source
AI summary
A method of determining fraud includes: receiving a transaction request associated with a first payment transaction between a merchant and a user from a merchant system; generating a first risk score based on the transaction request and a first set pot of transaction data received prior to the transaction request; processing a transaction request approval based on the first risk score not satisfying a first threshold; receiving a risk score request associated with the first payment transaction, where the risk score request is received after the transaction request has been approved; generating a second risk score based on a second set of transaction data received after the first risk score is determined; and automatically classifying the first payment transaction as potentially fraudulent in response to determining that the second risk score satisfies a second threshold.