数据与计算发展前沿 ›› 2026, Vol. 8 ›› Issue (2): 241-249.

CSTR: 32002.14.jfdc.CN10-1649/TP.2026.02.018

doi: 10.11871/jfdc.issn.2096-742X.2026.02.018

• 技术与应用 • 上一篇    下一篇

Transformer语言模型下无线网云数据伪造攻击自适应检测

王峰*()   

  1. 上海市浦东新区精神卫生中心， 同济大学附属精神卫生中心， 上海 200124
  • 收稿日期:2025-03-05 出版日期:2026-04-20 发布日期:2026-04-23
  • 通讯作者: *王峰（E-mail: un819723@163.com）
  • 作者简介:王峰，上海市浦东新区精神卫生中心同济大学附属精神卫生中心，工程师，主要研究方向为自然语言大模型学习技术，网络入侵自适应检测技术。
    本文负责文章框架拟定、数据收集、全文撰写。
    WANG Feng, Engineer at Shanghai Pudong New Area Mental Health Center, Tongji University School of Medicine. His main research interests include large language model learning technologies and adaptive network intrusion detection techniques.
    In this paper, he is mainly responsible for formulating the article framework, data collection, and manuscript writing.
    E-mail: un819723@163.com

Adaptive Detection of Wireless Cloud Data Falsification Attacks in Transformer Language Model

WANG Feng*()   

  1. Shanghai Pudong New Area Mental Health Center, Tongji University School of Medicine, Shanghai 200124, China
  • Received:2025-03-05 Online:2026-04-20 Published:2026-04-23

RichHTML

0

PDF

5

摘要：

【应用背景】 在无线网云数据应用过程中，根据单一邻域信息的注意力，整合的检测路径云数据序列较为单一，检测矩阵输出值具有单一性。【文献范围】在多步伪造攻击中，导致MSE值、MAE值、GAME值异常波动。【目的】 为满足自适应检测需求，设计了Transformer语言模型下的无线网云数据伪造攻击自适应检测方法。【方法】 分析可疑路径中多跳逆向邻域信息的注意力，将多个节点的邻域信息整合，得到无线网云数据多步伪造攻击检测路径。利用Transformer模型，将检测路径输入序列映射到多个子空间，每个子空间中应用独立的注意力机制，建立路径输入数据序列检测矩阵，并生成云数据攻击检测矩阵自适应语言学习字典，从而实现无线网云数据伪造攻击的精准检测。【结果】 实验结果表明：在聚合的检测路径中，文献[3]方法的检测MAE值最大值高于0.009，MSE值达到了0.09以上，GAME值达到了0.07以上，检测误差较大，而所提方法的MSE值仅在0.002以内变化，MAE值在0.02以内变化，GAME值在0.01以内变化，伪造攻击自适应检测误差更小，攻击检测准确性更高。【局限】该方法需要高性能硬件支持，同时具备跨平台兼容性。【结论】 伪造攻击自适应检测准确性较高，对于提升无线网数据安全性具有重要作用。

关键词: Transformer语言模型, 无线网, 云数据, 伪造攻击, 自适应检测方法

Abstract:

[Application Background] In the process of wireless network cloud data application, the integrated detection path cloud data sequence is relatively single based on the attention of a single neighborhood information, and the output value of the detection matrix has singularity. [Scope of Literature] Under multi-step forgery attacks, existing methods may cause abnormal fluctuations in the MSE, MAE, and GAME values [Purpose] To meet the requirements of adaptive detection, this paper proposes an adaptive detection method for wireless network cloud data forgery attacks based on a Transformer model. [Method] The proposed method analyzes the attention to multi-hop reverse neighborhood information in suspicious paths and integrates neighborhood information from multiple nodes to construct a multi-step forgery attack detection path for wireless network cloud data. Based on the Transformer model, the input sequence of the detection path is mapped into multiple subspaces, where independent attention mechanisms are applied in each subspace. A detection matrix for the path input data sequence is then established, and an adaptive language learning dictionary for the cloud data attack detection matrix is generated, thereby enabling accurate detection of wireless network cloud data forgery attacks. [Result] Experimental results show that, in the aggregated detection path, the maximum MAE value of the method proposed in reference [3] is higher than 0.009, the MSE value reaches above 0.09, and the GAME value reaches above 0.07, indicating a large detection error. However, the MSE value of the proposed method only changes within 0.002, the MAE value changes within 0.02, and the GAME value changes within 0.01, indicating smaller adaptive detection errors and higher detection accuracy for forgery attacks. [Limitations] This method requires high-performance hardware support and cross platform compatibility. [Conclusion] The adaptive detection accuracy of forgery attacks is high, which plays an important role in improving the security of wireless network data.

Key words: Transformer language model, wireless network, cloud data, fake attacks, adaptive detection method