Graph Neural Network Scam Detection via Structural Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine-learning algorithms for scam website detection rely on fixed models and assume unchanged attack vectors, making them ineffective for detecting new or unknown scam websites.
Innovation Solution
Utilize graph neural networks (GNNs) to create a dataset of target websites, extract website constructs, build a graph associating websites with scam categories, and train an inductive GNN to identify potential scam websites based on similarity scores without model retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional machine-learning algorithms with fixed models are used for scam website detection, then the detection process is simple and fast, but the detection effectiveness deteriorates when encountering new or unknown scam websites with changed attack vectors
Solution Approach 1:
The patent transforms the static fixed model into a dynamic system by implementing continuous learning through feedback loops. The system automatically retrains models with newly identified scam patterns, allowing the detection algorithm to adapt and evolve alongside changing scam tactics, thereby maintaining high detection effectiveness against novel threats
Solution Approach 2:
The system incorporates feedback mechanisms where detection results and newly identified scam websites are fed back into the model retraining process. This closed-loop approach enables the system to learn from actual scam patterns encountered in practice, continuously improving its ability to detect both known and emerging scam types
2Reliability
If model retraining is implemented to improve detection of new scam websites, then detection effectiveness improves, but system complexity and computational resources increase
Solution Approach 1:
The system implements self-service automation where the model retraining process is triggered automatically based on detected scam patterns without requiring manual intervention. The system autonomously collects data, retrains models, and deploys updated detection algorithms, reducing operational complexity despite the enhanced detection capabilities
Solution Approach 2:
The system performs preliminary actions by pre-processing and storing scam website data in structured formats beforehand. This preparation work is done in advance so that when retraining is needed, the system can quickly utilize the pre-organized data without extensive processing delays, managing computational complexity more efficiently
3Measurement precision
If feature set is expanded to capture diverse scam characteristics, then detection precision improves, but model retraining becomes necessary increasing time consumption
Solution Approach 1:
The system applies partial retraining by selectively updating only those model components affected by new scam patterns rather than retraining the entire model from scratch. This approach maintains detection precision for new threats while significantly reducing the time and computational resources required compared to full model retraining
Data Source
AI summary
A computer-implemented method for utilizing graph neural networks for scam website detection may include (i) creating a dataset of target websites including unknown websites and known scam websites with corresponding scam categories, (ii) extracting website constructs utilized for executing scam attacks from the dataset of target websites, (iii) building a graph including nodes for associating each of the target websites with one or more of the corresponding scam categories, (iv) grouping, for each of the nodes, the target websites based on sharing a common construct within the website constructs, and (v) performing a security action that trains a graph neural network for identifying the unknown websites as potential scam websites based on a similarity score determined from the website constructs. Various other methods, systems, and computer-readable media are also disclosed.