Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 41 of the 49 references it cites were found in a registry or an index; 9 of the 13 sources its claims cite were checked for support.

HuntGPT: Integrating Machine Learning-Based Anomaly Detection and Explainable AI with Large Language Models (LLMs)

Open access

Abstract

Machine learning (ML) methods for network anomaly detection are emerging as effective proactive strategies in threat hunting, substantially reducing the time required for threat detection and response. However, the challenges in training and maintaining ML models, coupled with frequent false positives, diminish their acceptance and trustworthiness. In response, Explainable AI (XAI) techniques have been introduced to enable cybersecurity operations teams to assess alerts generated by AI systems more confidently. Despite these advancements, XAI tools have encountered limited acceptance from incident responders and have struggled to meet the decision-making needs of both analysts and model maintainers. Large Language Models (LLMs) offer a unique approach to tackling these challenges. Through tuning, LLMs have the ability to discern patterns across vast amounts of information and meet varying functional requirements. In this research, we introduce the development of HuntGPT, a specialized intrusion detection dashboard created to implement a Random Forest classifier trained utilizing the KDD99 dataset. The tool incorporates XAI frameworks like SHAP and Lime, enhancing user-friendliness and intuitiveness of the model. When combined with a GPT-3.5 Turbo conversational agent, HuntGPT aims to deliver detected threats in an easily explainable format, emphasizing user understanding and offering a smooth interactive experience. We investigate the system’s comprehensive architecture and its diverse components, assess the prototype’s technical accuracy using the Certified Information Security Manager (CISM) Practice Exams, and analyze the quality of response readability across six unique metrics. Our results indicate that conversational agents, underpinned by LLM technology and integrated with XAI, can enable a robust mechanism for generating explainable and actionable AI solutions, especially within the realm of intrusion detection systems.

Overview

CoChat Summary

TL;DR

HuntGPT is an intrusion-detection dashboard that combines a Random Forest classifier trained on KDD99, SHAP/LIME explanations, and a GPT-3.5 Turbo conversational agent to make network anomaly alerts more explainable and actionable. Evaluation shows GPT-3.5 Turbo achieved 72–82.5% success on cybersecurity certification practice exams, while generated explanations and chatbot answers were mostly at graduate-level readability.

Key findings

  • GPT-3.5 Turbo’s cybersecurity knowledge was tested on three practice exams: CISM Certified Information Security Manager Practice Exams (40 questions, 82.5% success), ISACA official CISM practice quiz (10 questions, 80%), and ISACA official cybersecurity fundamentals practice quiz (25 questions, 72%).
  • Readability analysis of 20 generated anomaly explanations and 20 chat logs found mostly graduate-level text: Flesch-Kincaid Grade Level 15.7 for explanations and 14.9 for chatbot answers; Flesch Reading Ease 22.7 and 23.9; Dale Chall 12.6 and 12.11; Automated Readability Index 16.3 and 15.9; Coleman Liau Index 15.3 and 16.3; Linsear Write 17.4 and 15.8.
  • The paper reports that the chatbot produced actionable responses, with example queries about preventing DoS attacks, firewalls, free firewall options, installing ZoneAlarm, and configuring a small home network.
  • The prototype is presented as integrating ML anomaly detection, XAI, and LLM-based conversation to improve analyst understanding and trust in intrusion detection.

Methods

  • Design: HuntGPT is organized into three layers—an analytics engine for packet analysis, Elasticsearch/S3 for data storage, and a Gradio dashboard UI integrated with OpenAI’s Language Model API.
  • Detection and explanation: a Random Forest classifier trained on the KDD99 dataset classifies network flows; SHAP and LIME generate explanations, plots, and JSON documents stored in Elasticsearch.
  • Evaluation: technical accuracy was assessed using CISM and ISACA practice exams; response quality was assessed on 20 generated explanations and 20 chat logs, each with approximately six questions, using six readability formulas from the Python py-readability-metrics library.

Limitations

  • The paper states future work is needed to refine the ML model and incorporate real-time detection capabilities, and to embed actionable AI so the chatbot can issue active commands directly to a SIEM component.
  • Evident from the text: evaluation relies on certification practice exams and readability metrics rather than operational SOC deployment, so real-world false-positive reduction, analyst decision accuracy, and production performance are not measured.
  • Evident from the text: the response-quality evaluation uses a small sample of 20 generated explanations and 20 chat logs, and the KDD99 dataset is an older benchmark, which may limit generalization to current network traffic.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/telecom7030073

Cites

41 of 49 references verified0 citation errors · 8 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

7 Valid · 2 Not valid · 3 Not checked · 18 Own results and interpretations

6 more claims

Create a free account to see all 30 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 16
  • It may state two separate claims as one. 13

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Cited by

See all 104

Metadata

Authors
Ali, T., Kostakos, P., & Sheikhi, S.
Affiliations
University of Oulu (FI)
Venue
Telecom
Publisher
MDPI AG
Published
2026-06-08
Volume
7
Issue
3
Pages
73
ISSN
2673-4001 (ISSN-L)
OpenAlex
W4387210439
Cited by
104
References
48
Citations by year
2026: 33
FWCI
91.18
Type
journal article
Language
en
Open access
Locations
  • Telecom (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • arXiv (Cornell University) (cc-by)
Funders
European Commission (101021911, 318927)
Topics
  • Physical Sciences › Computer Science › Computer Networks and Communications › Network Security and Intrusion Detection
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
Keywords
network anomaly detection, threat hunting, explainable AI, large language models, intrusion detection systems, random forest, SHAP, lime, GPT-3.5-turbo, chatbots, KDD Cup 1999 dataset, false positives, incident response

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-07 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.