Deep-Learning OAO Fraud Detection Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems face challenges in detecting fraudulent online account origination (OAO) applications, particularly in distinguishing between human and human-like behavior, which complicates authenticating new applicants and managing numerous form types, leading to redundant models and inefficiencies in fraud detection.
Innovation Solution
The system employs machine learning using meta-learning and Bayesian learning to reduce the number of models per form type by half and provides a quick start for new applicants by selecting the best existing model based on available data, utilizing deep-learning to differentiate between normal and nefarious user behaviors during online application submissions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If one model is created per form type, then the system can accurately detect fraud for each specific form type, but the device complexity increases and becomes difficult to manage as the number of form types grows
Solution Approach 1:
The patent applies universality by training a single OAO model on aggregated data from multiple form types, enabling one model to perform fraud detection across diverse form types. The model learns universal fraudulent behavior patterns that generalize across different forms, eliminating the need for separate models for each form type while maintaining detection effectiveness.
Solution Approach 2:
The patent merges data from multiple form types into a unified training dataset, combining information across different forms to train a single comprehensive model. This aggregation approach consolidates what would otherwise require multiple separate models into one unified system, reducing complexity while preserving detection capability.
2Adaptability or versatility
If a new model is trained for each new client, then the system can adapt to client-specific fraud patterns, but the loss of time increases due to extended training requirements
Solution Approach 1:
The patent applies preliminary action by pre-training the OAO model on aggregated data from multiple clients before deployment. This advance training establishes a robust baseline model that can immediately handle new clients without requiring time-consuming training sessions, enabling rapid onboarding while maintaining adaptability through the model's learned patterns.
Solution Approach 2:
The patent uses copying by replicating successful training approaches from existing clients to new clients. The model learns from aggregated data across clients and applies these learned patterns to new clients, effectively copying proven detection strategies without requiring de novo training for each client.
3Reliability
If traditional fraud detection methods are used, then the system can identify obvious fraudulent applications, but the difficulty of detecting and measuring increases when applications appear human or human-like
Solution Approach 1:
The patent replaces traditional rule-based mechanical fraud detection systems with a machine learning-based OAO model. This substitution enables the system to automatically learn and detect subtle behavioral patterns that indicate fraud, even when applications appear human-like, overcoming the limitations of conventional detection methods through adaptive pattern recognition.
Data Source
AI summary
A fraud prevention server that includes an electronic processor and a memory. The memory includes an online application origination (OAO) service and a plurality of OAO models, each of the plurality of OAO models differentiates between a behavior of a normal user and a behavior of a nefarious actor during a submission of the online application on a device. When executing the OAO service, the electronic processor is configured to receive form data from a client server, determine a best OAO model from a plurality of OAO models with deep-learning, determine a fraud score of the online application based on the best OAO model, and control the client server to approve, hold, or deny the online application based on the fraud score that is determined.