Prompt word attack detection method and system of large language model and electronic equipment

By employing a three-tiered, progressive detection method, combined with heuristic semantic parsing, a Transformer-based language model, and structured verification of a large language model, the problem of insufficient detection accuracy and long text processing capabilities of large language models for prompt word injection attacks is solved, achieving efficient and accurate identification and real-time response to complex attacks.

CN121389113AActive Publication Date: 2026-01-23HANG ZHOU LING XIN SHU KE XIN XI JI SHU YOU XIAN GONG SI
4 Cites 5 Cited by

Patent Information

Application Number
CN202511984007.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-12-26
Publication Date
2026-01-23
Estimated Expiration
2045-12-26

AI Technical Summary

Technical Problem

Existing technologies suffer from low detection accuracy and insufficient ability to handle long texts when detecting prompt injection attacks on large language models. In particular, the false negative rate is high in complex attack scenarios, and a single method cannot effectively balance detection accuracy and computational performance.

Method used

A three-tiered detection method is adopted. First, obvious suspicious text blocks are quickly screened out through heuristic semantic parsing. Then, deep semantic analysis is performed through a first language model based on the Transformer architecture. Finally, a large language model is used to perform structured verification of the attack intent, and meta-clue words are combined for final adjudication.

Benefits of technology

It achieves efficient and accurate detection of complex attacks, takes into account the needs of long text processing, reduces computational latency, improves the real-time performance and detection accuracy of the system, and forms a defense-in-depth system to ensure the identification and interpretability of new and complex attacks.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention provides a cue word attack detection method and system of a large language model and electronic equipment, and relates to the technical field of text process.The method comprises the steps that a to-be-processed text block in an original text is determined; performing heuristic semantic analysis on the text block to obtain a first suspicious text block and a first risk score thereof; performing semantic analysis on the first suspicious text block and the context information thereof through a first language model to obtain a second suspicious text block and a second risk score thereof; performing structured verification of the attack intention on the second suspicious text block according to a preset meta prompt word through a second language model to obtain an attack intention verification result; the attack intention verification result comprises a third suspicious text block and a third risk score thereof; and executing a target response action on the original text based on the first risk score, the second risk score and the third risk score. According to the method, the efficient processing requirement of the long text can be considered, and the detection accuracy of complex attacks can be improved.
Need to check novelty before this filing date? Find Prior Art