Real-Time Call Fraud Detection Using Sliding Window Transcripts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting telecommunication fraud are often inaccurate and may lead to missed or incorrect or determinations, leaving the user vulnerable to fraudulent activity, particularly in cases where fraudsters change their phone numbers through VoIP, and existing systems fail to accurately identify fraudulent calls in real-time.
Innovation Solution
A computer-implemented method using sliding window segmentation for real-time transcription of calls, determining window and aggregate probabilities of fraud, and sending alerts when the aggregate probability exceeds a threshold, with features like speaker identification and adaptive weightings to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional blacklist methods are used to detect fraud, then implementation is simple, but accuracy is low and fraudsters can easily evade detection by changing phone numbers
Solution Approach 1:
The patent segments the telecommunication call into multiple time windows and analyzes each window separately. The call is divided into overlapping time segments, with each segment being analyzed for fraud indicators independently, then aggregated to form an overall fraud probability assessment. This segmentation allows the system to capture fraud patterns at different stages of the call while maintaining computational efficiency.
Solution Approach 2:
The patent implements dynamic fraud probability calculation by continuously updating the aggregate probability as new time windows are analyzed. The system transitions from static blacklist matching to dynamic real-time probability assessment, where the fraud determination evolves throughout the call duration based on accumulating evidence from multiple time windows.
2Speed
If real-time fraud detection is implemented, then user protection is immediate, but computational accuracy may be reduced due to limited processing time
Solution Approach 1:
The patent performs preliminary fraud indicator extraction and window probability calculation for each time window before aggregating them. By pre-processing and analyzing individual windows in real-time, the system prepares fraud evidence incrementally throughout the call, enabling rapid final aggregation and decision-making without requiring complete call analysis beforehand.
Solution Approach 2:
The patent maintains continuous fraud detection throughout the call by analyzing overlapping time windows sequentially. The detection process continues uninterrupted from call start to finish, with each window contributing to the cumulative fraud probability. This continuous analysis ensures real-time protection while accumulating sufficient evidence for accurate determination.
3Measurement precision
If sliding window segmentation is used for real-time analysis, then detection accuracy improves through context consideration, but computational complexity increases
Solution Approach 1:
The patent divides the call transcript into multiple time windows with overlap, where each window is analyzed for fraud indicators. The segmentation creates manageable analysis units that can be processed independently and then aggregated, reducing the computational burden of analyzing the entire call at once while maintaining context through overlapping windows.
Solution Approach 2:
The patent analyzes only the necessary portions of the call at each time window rather than processing the entire call continuously. By focusing computation on specific time segments with overlap, the system achieves sufficient detection precision without the excessive computational cost of analyzing every moment of the call with full context.
Data Source
Figure 1
Figure 2A~2B
Figure 2C
AI summary
A computer-implemented method and system is disclosed for detecting telecommunication fraud. The method comprises obtaining a real-time transcription of a telecommunication call during the call and segmenting, in real-time, the transcription using sliding window segmentation. The method then determines, for a leading window, a window probability of telecommunication fraud. The method further comprises generating, in real-time, an aggregate probability of telecommunication fraud based on the window probability of the leading window and a window probability of at least one previous window, and sending a signal to generate an alert in the event that the aggregate probability exceeds a threshold.