Distributed Fraud Detection Model Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current fraud detection systems are inefficient in identifying fraudulent transactions, as they rely on centralized models that blacklist transactions based on known fraudulent locations, leading to significant financial losses despite sophisticated detection systems.
Innovation Solution
Distributing a fraud detection model to customer devices, where customers are grouped by common characteristics and transactions are voted on by multiple models within clusters to classify transactions as fraudulent or non-fraudulent, leveraging the processing power of millions of devices for robust detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a centralized fraud detection model is used, then the system complexity is reduced, but the detection accuracy and robustness deteriorate due to reliance on blacklisted locations and insufficient diverse training data
Solution Approach 1:
The patent divides the centralized fraud detection model into multiple distributed instances, each deployed on customer devices. These segmented model instances process transactions independently using local data, eliminating the need for a single centralized model while improving detection accuracy through diverse training data from multiple customers.
Solution Approach 2:
The patent creates multiple copies of the fraud detection model and distributes them to different customer devices. Each copy is trained on local transaction data from its host customer, enabling the system to leverage diverse training data across multiple instances without requiring a single complex centralized model.
2Device complexity
If a centralized fraud detection model is used, then model distribution is simplified, but the computational resources required for training and processing deteriorate due to handling all transactions through a single model
Solution Approach 1:
The patent segments the computational workload by distributing model instances across multiple customer devices. Each device handles only its local transactions, reducing the computational burden on any single model while collectively processing a larger volume of transactions efficiently.
Solution Approach 2:
The patent applies partial action by having each distributed model instance process only the transactions relevant to its host customer rather than all transactions. This partial processing approach reduces overall computational resources required compared to a centralized model that must evaluate every transaction.
3Reliability
If transactions are evaluated by multiple distributed models, then detection robustness improves through diverse training data, but the processing time and system response time deteriorate
Solution Approach 1:
The patent uses partial action by having each distributed model instance perform rapid local evaluation of transactions independently. Multiple models process transactions in parallel rather than sequentially, reducing total processing time while maintaining robustness through diverse perspectives from different trained models.
Solution Approach 2:
The patent enables self-service by allowing each distributed model instance to autonomously evaluate transactions using its own trained parameters without requiring coordination or sequential processing. This independent self-evaluation capability significantly reduces processing time compared to centralized sequential processing.
4Reliability
If distributed fraud detection models are implemented, then detection accuracy and robustness improve through diverse training data, but the device complexity and system architecture deteriorate
Solution Approach 1:
The patent applies self-service by implementing autonomous model instances on customer devices that independently train and evaluate transactions without requiring complex centralized coordination. Each model serves itself by using local data for training and decision-making, simplifying the overall system architecture despite the distributed nature.
Solution Approach 2:
The patent uses parameter changes by adapting the fraud detection model to different local conditions and training data characteristics at each customer device. This allows the system to maintain high detection accuracy through localized parameter optimization while avoiding complex centralized management by letting each instance adapt independently.
Data Source
AI summary
Logic may assign a customer identification to a model to associate a first customer with the model to detect fraudulent transactions. Logic may determine one or more clusters to associate with the first customer based on characteristics associated with the first customer. Logic may associate one or more cluster identifications with the first customer. Each cluster identification may identify one cluster of the one or more clusters. Each cluster may identify a group of customers based on characteristics associated with the group of customers. Logic may cause the model to transmit to a customer device associated with the first customer. Logic may receive transaction data for a transaction for one customer of the group of customers associated with a first cluster. And logic may communicate modified transaction data to customer devices of more than one customer of the group of customers associated with the first cluster.