Multi-Level Prompt Injection Detection Framework Using Attention Pattern Analysis 


Vol. 15,  No. 4, pp. 281-289, Apr.  2026
https://doi.org/10.3745/TKIPS.2026.15.4.281


PDF
  Abstract

As the utilization of Large Language Models (LLMs) expands, prompt injection attacks, which manipulate user inputs to bypass model security policies, have emerged as a significant threat. Existing defense techniques have faced limitations in actual service environments due to the trade-off between the low accuracy of rule-based methods and the high computational costs of model-based detection. To address these challenges, this study proposes a high-efficiency multi-level de4fense framework that combines Banned Terms Filtering with Attention Pattern Analysis (Attention Tracker). The proposed system minimizes system load by filtering explicit attack keywords in the first stage and identifies sophisticated attacks that manipulate context by analyzing changes in the LLM’s attention weights using the Focus Score in the second stage. Experimental results using a dataset of actual malicious prompts demonstrate that the proposed method reduces latency by over 50% and lowers computational costs compared to existing single-model approaches, while achieving approximately a two-fold increase in detection accuracy (83%). This study is significant in that it presents a practical security detection sy stem capable of stable operation in LLM service environments by effectively securing both computational efficiency and detection performance.

  Statistics
  Cite this article

[IEEE Style]

J. S. Woo, K. N. Ryeong, L. I. Gu, "Multi-Level Prompt Injection Detection Framework Using Attention Pattern Analysis," The Transactions of the Korea Information Processing Society, vol. 15, no. 4, pp. 281-289, 2026. DOI: https://doi.org/10.3745/TKIPS.2026.15.4.281.

[ACM Style]

Jeong Sun Woo, Kim Nam Ryeong, and Lee Il Gu. 2026. Multi-Level Prompt Injection Detection Framework Using Attention Pattern Analysis. The Transactions of the Korea Information Processing Society, 15, 4, (2026), 281-289. DOI: https://doi.org/10.3745/TKIPS.2026.15.4.281.

Forms

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)

Old Journals

Indexing

All publications of TKIPS is indexed in DOI, EBSCO, Google Scholar, Crossref, and CrossCheck.

TKIPS is also selected as the Journal for Accreditation by NRF (National Research Foundation of Korea).

Related Journals