Contrastive Sequence Models for Card Testing Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models for fraud detection face challenges in adapting to evolving fraud patterns, lack flexibility, and struggle with insufficient labeled data, especially for emerging fraud schemes like card testing, leading to inefficiencies and missed contextual insights.
Innovation Solution
A sequence-based contrastive learning model is employed to detect fraud, utilizing a first computer model trained on unlabeled data with contrastive techniques like Triplet Loss and Info NCE loss, and a second model to enhance training, focusing on transaction sequences and patterns, enabling efficient detection of card testing activity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional supervised learning techniques are used to train machine learning models for fraud detection, then the model can be trained to identify fraudulent transactions, but the process becomes time-consuming and resource-intensive due to the difficulty of acquiring labeled data
Solution Approach 1:
The patent introduces an intermediary contrastive learning model that acts as a mediator between the unlabeled transaction data and the fraud detection task. This intermediate model learns meaningful representations from unlabeled data through contrastive loss functions, which then serve as training data for the final fraud detection model, thereby eliminating the time-consuming manual labeling process while maintaining detection accuracy
Solution Approach 2:
The patent creates synthetic labeled data by copying and transforming unlabeled transaction data through contrastive learning. The contrastive learning model generates pseudo-labels and transformed data representations that replicate the structure and patterns of actual fraudulent transactions, allowing the fraud detection model to be trained on these synthesized copies rather than requiring genuine labeled examples
2Adaptability or versatility
If conventional machine learning models use pre-defined workflows to access data from a single electronic data source, then the workflow is simple and straightforward, but the model lacks flexibility and cannot adapt well to evolving fraud patterns or specific user needs
Solution Approach 1:
The patent implements a dynamic data access mechanism where the machine learning model can flexibly select and access relevant data sources and features based on the specific characteristics of each transaction. The system dynamically adjusts its data retrieval workflow according to the transaction type, amount, and other attributes, enabling adaptation to evolving fraud patterns without requiring complex manual reconfiguration
Solution Approach 2:
The patent creates a universal machine learning model that can handle multiple types of transactions and fraud patterns through a single integrated architecture. The model incorporates multiple data sources (transaction data, device information, network data) and can adapt its analysis approach based on the specific transaction being evaluated, providing both simplicity and versatility
3Loss of information
If machine learning models focus solely on individual transactions, then the analysis is simple and direct, but the model misses contextual insights that can be gleaned from analyzing sequences of transactions associated with a specific computing system
Solution Approach 1:
The patent segments the transaction analysis into two complementary levels: individual transaction analysis and sequence-level contextual analysis. The model first evaluates each transaction independently for basic fraud indicators, then analyzes sequences of transactions from the same computing system to identify contextual patterns and anomalies that would be invisible at the individual transaction level
Solution Approach 2:
The patent adds a temporal and contextual dimension to the fraud analysis by examining sequences of transactions over time. Instead of analyzing transactions in isolation, the model incorporates time-based patterns, repeated behaviors, and contextual relationships between transactions from the same device, thereby uncovering hidden fraud patterns that require this additional dimensional perspective
Data Source
AI summary
Discussed herein are methods and systems to train customized machine learning models in a more efficient manner (e.g., using fewer labeled data points). In one example, a method may include using a first machine learning to generate likelihoods of fraudulent activity for an aggregated series of data associated with a series of computing systems. Based on the calculated likelihoods, a server can generate a training dataset that includes fraudulent data associated with a first computing system, fraudulent data associated with any other computing system within the series of computing systems other than the first computing system, non-fraudulent data associated with the first computing system, and non-fraudulent data associated with any other computing system within the series of computing systems other than the first computing system. The server may then train a second machine learning model using the training data, e.g., using a contrastive learning method.