Cross-Service Fraud Estimation Using Fraud and Authentic User Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fraud estimation systems struggle to accurately identify fraudulent users whose activities do not resemble those on a blacklist, limiting the precision of fraud detection.
Innovation Solution
A fraud estimation system that utilizes a learning model to analyze user information across multiple services, comparing user data from different platforms to determine fraud likelihood, incorporating utilization situations, and employing machine learning techniques like neural networking and deep learning to enhance precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If comparison is made only with users on the blacklist, then users with similar fraudulent patterns can be detected, but fraud detection precision cannot be raised to a satisfactorily high level because fraudulent users' activities do not always resemble those on the blacklist
Solution Approach 1:
The patent segments the comparison target into two distinct groups: fraudulent users and authentic users. By separately comparing the target user's action history against both groups, the system can identify patterns that distinguish fraudulent behavior from legitimate behavior, thereby improving detection precision for diverse fraudulent patterns that may not resemble existing blacklist entries.
Solution Approach 2:
The patent changes the comparison parameter from solely matching fraudulent patterns (blacklist similarity) to a dual-parameter approach that measures both similarity to fraudulent users and similarity to authentic users. This parameter transformation enables the system to detect fraud even when the fraudulent pattern is novel or different from known blacklist patterns, by identifying deviations from authentic user behavior.
2Ease of manufacture
If action histories are compared to blacklist users only, then detection can be performed with simple methodology, but only users whose activities resemble those on the blacklist are detectable
Solution Approach 1:
The patent divides the reference user group into two segments: fraudulent users and authentic users. This segmentation maintains methodological simplicity by using the same comparison technique for both groups while significantly improving detection reliability. The system can now reliably detect both known fraudulent patterns and novel fraudulent patterns that differ from authentic behavior, without complicating the core comparison methodology.
3Adaptability or versatility
If multiple services are integrated for comparison, then fraud detection coverage is improved, but system complexity increases
Solution Approach 1:
The patent implements a universal comparison framework that can handle multiple services through a single unified mechanism. The action history comparison method is designed to be service-agnostic, accepting action histories from any service and comparing them against reference users from any service. This multi-functional approach improves cross-service fraud detection coverage while avoiding the need for service-specific detection systems, thereby limiting the increase in overall system complexity.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Storage means (100c) of a fraud estimation system (S) stores a learning model that has learned a relationship between a comparison result that is a result of comparing user information of a user in one service to user information of a fraudulent user or an authentic user in another service and presence or absence of fraudulence in the one service. Comparison result obtaining means (104c) obtains a comparison result that is a result of comparing user information of a target user in the one service and user information of a fraudulent user or an authentic user in the another service. Output obtaining means (105c) obtains output from the learning model based on the comparison result. Estimation means (106c) estimates fraudulence of the target user based on the output from the learning model.