A phishing account detection model training method, detection method and device

CN121365975BActive Publication Date: 2026-07-14BEIJING UNIV OF POSTS & TELECOMM

Patent Information

Authority / Receiving Office: CN · China
Patent Type: Patents(China)
Current Assignee / Owner: BEIJING UNIV OF POSTS & TELECOMM
Filing Date: 2025-09-29
Publication Date: 2026-07-14

AI Technical Summary

Technical Problem

Existing blockchain phishing detection technologies struggle to identify phishing accounts that use small-amount transfers and visually obfuscated addresses in real time and accurately. This is especially true on large decentralized exchanges, where traditional methods are computationally expensive and carry a high risk of misjudgment.

Method used

By constructing a second-order transaction network independently for each account, using a random walk restart algorithm to generate local subgraphs, combining graph neural networks and classifiers, introducing address similarity features, and constructing node-level contrast loss and classification loss, lightweight incremental update phishing account detection is achieved.

Benefits of technology

Significantly reduces the risk of false positives in large-scale decentralized trading platforms, achieving highly sensitive and real-time phishing account detection, suitable for online risk control scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

Figure CN121365975B_ABST

Patent Text Reader

Abstract

The application provides a fishing account detection model training method, a detection method and a device, a second-order transaction network is independently constructed for each account, the local transaction structure is completely retained, the feature dilution and excessive smoothing bottleneck caused by the traditional global large graph training are broken through, an address similarity feature is introduced, the high-frequency transaction object tail number similarity deliberately constructed by an attacker is accurately captured, and visual confusion type phishing behavior is effectively identified. A random walk restart algorithm is used to generate a double local subgraph, a node level contrast loss is constructed by taking the self-pairing of a center node as a positive sample and a cross-subgraph non-center node as a negative sample, and the discrimination ability of the model for a phishing node and a normal neighbor is strengthened; an end-to-end training is performed in combination with a classification loss, the graph neural network can be incrementally updated without retraining a global graph, and lightweight deployment is realized. In the blockchain transaction scene with extremely unbalanced data, the application can significantly reduce the misjudgment risk of a phishing account, and has high sensitivity and real-time performance.

Need to check novelty before this filing date? Find Prior Art