Introduction
Telecommunication fraud remains a persistent global threat, costing billions annually. While traditional fraud detection systems are effective against persistent criminal patterns, they often fail to identify "mutation fraud." In this scenario, fraudsters acquire or rent SIM cards from legitimate users, maintaining a facade of normal behavior for extended periods before executing a sudden burst of fraudulent activity. This subtle shift makes mutation fraud exceptionally difficult to detect using standard statistical averages, which tend to smooth over these critical, short-term behavioral changes.
The MutationGuard Framework
Presented at IJCAI-25, MutationGuard is a deep learning architecture specifically engineered to capture the local and temporal nuances of mutation fraud. Unlike previous methods that rely on manual feature engineering, MutationGuard processes raw call records and social graph data through a dual-encoder system.
The framework consists of two primary components:
- User Encoder: Utilizes a Graph Attention Network (GAT) to monitor changes in a user's communication network. Legitimate users typically maintain dense connections with known contacts, whereas mutation fraudsters exhibit a sudden shift toward numerous "one-off" calls to unconnected strangers.
- VOC (Voice of Call) Encoder: Analyzes the temporal and spatial patterns of individual call records.
Mapping Time to 3D Geometry
The core innovation of MutationGuard lies in its treatment of time. Traditional 1D timelines fail to capture the periodic nature of human behavior. To address this, the researchers developed Spatial Position Calculation (SPC), which maps call records onto a 3D cylindrical spiral.
In this representation, one full rotation of the spiral corresponds to 24 hours. This geometric approach ensures that calls made at the same time on different days are spatially adjacent, allowing the model to apply point-cloud processing techniques (like PointNet++) to identify clusters of suspicious activity. This is complemented by a Temporal Change Feature Extraction (TCFE) module using BiLSTM to track how these behavioral clusters evolve over time.
Evaluation and Performance
The model was rigorously tested on two datasets: the newly curated MutationTeleFraud dataset (containing records for 9,874 subscribers) and the public Sichuan Telecommunication Fraud Dataset.
MutationGuard demonstrated superior performance across all metrics:
- Accuracy: Improved detection of mutation fraud cases by 13.93% over the strongest baseline.
- Robustness: Achieved an AUC of 0.9898 and an F1 score of 0.9396.
- Scalability: The system is capable of processing over 250,000 users per hour on a single Tesla V100 GPU, proving its readiness for large-scale industrial deployment.
Conclusion
By rethinking temporal data as 3D geometry and combining it with graph-based social analysis, MutationGuard provides a powerful defense against the most deceptive forms of telecom fraud. Its ability to learn complex patterns directly from raw data represents a significant advancement over rule-based systems, offering a scalable and highly accurate solution for modern telecommunication networks.
Summary updated