AutoDP-LLM: Automating Data Pre-processing for Intrusion Detection Systems using Large Language Models

Bao-Phong Nguyen; Gia-Khanh Pham; Thai-Duong Do; Mai Xuan Trang; Minh-Tuan Le; Xuan-Nam Tran; Huan Vu; Tien-Cuong Nguyen

cs.CROct 4, 2026

AutoDP-LLM: Automating Data Pre-processing for Intrusion Detection Systems using Large Language Models

Authors: Bao-Phong Nguyen, Gia-Khanh Pham, Thai-Duong Do, Mai Xuan Trang, Minh-Tuan Le, Xuan-Nam Tran, Huan Vu, Tien-Cuong Nguyen, +2 more

Organizations: Business AI Lab, College of Technology, National Economics University, Hanoi, Vietnam

Abstract

The increasing complexity and scale of modern cyber-attacks demand intelligent and computationally efficient Intrusion Detection Systems (IDS). However, designing effective data pre-processing pipelines traditionally involves substantial trial-and-error effort and repeated evaluation of alternative configurations. For large, high-dimensional network traffic data, this process can create a significant computational burden. In this work, we propose AutoDP-LLM, an automated pre-processing framework designed to reduce manual pipeline development and computational overhead. Specifically, AutoDP-LLM leverages Large Language Models (LLMs) to autonomously generate and validate executable data pre-processing pipelines. The framework combines deterministic host-side planning with LLM-based specialist agents to formulate data-processing strategies, synthesize executable code, and adaptively determine retained feature sets using semantic reasoning and training-derived statistical evidence, without requiring a predefined feature budget. Focusing on multiclass intrusion detection, we evaluate AutoDP-LLM on the UNSW-NB15 and NSL-KDD benchmark datasets using multiple downstream classifiers. Comparative experiments against conventional feature-selection methods show that AutoDP-LLM achieves competitive detection performance while automating the generation of compact and executable pre-processing pipelines. Component-level ablation experiments further demonstrate the complementary contributions of the semantic and statistical feature-reduction components. The repeated generation, validation, execution, and assessment of candidate pipelines are amenable to parallel execution, highlighting the potential of scalable computing environments, including high-performance computing (HPC) systems, to support automated IDS pipeline development.

Figures & tables

Fig. 1: Overall pipeline of the IDS integrating our proposed AutoDP-LLM framework.

Fig. 2: Three-stage architecture and execution workflow of the proposed AutoDP-LLM framework.

Classifier	Accuracy	Precision	Recall	$F_{1}$ -score	Execution Time
Random Forest	$76.59\pm 0.28$	$82.64\pm 0.33$	$76.59\pm 0.28$	$78.41\pm 0.21$	$117.18\pm 9.80$
KNN	$73.33\pm 0.17$	$80.79\pm 0.33$	$73.33\pm 0.17$	$76.08\pm 0.18$	$108.98\pm 9.77$
MLP	77.51 $\pm$ 0.64	$81.17\pm 0.84$	77.51 $\pm$ 0.64	$77.76\pm 0.49$	$190.61\pm 9.30$
XGBoost	$77.20\pm 0.23$	82.84 $\pm$ 0.83	$77.20\pm 0.23$	78.41 $\pm$ 0.22	106.38 $\pm$ 9.77
LightGBM	$77.35\pm 0.30$	$82.82\pm 0.79$	$77.35\pm 0.30$	$78.40\pm 0.31$	$117.91\pm 9.72$
CatBoost	$76.84\pm 0.12$	$82.46\pm 0.76$	$76.84\pm 0.12$	$77.92\pm 0.14$	$150.56\pm 9.55$

TABLE I: Classification performance and total execution time on UNSW-NB15 across 30 pipeline-generation runs

Classifier	Accuracy	Precision	Recall	$F_{1}$ -score	Execution Time
Random Forest	$76.46\pm 0.85$	$81.97\pm 1.00$	$76.46\pm 0.85$	$72.43\pm 0.93$	$93.94\pm 14.59$
KNN	$77.22\pm 0.52$	$80.77\pm 0.72$	$77.22\pm 0.52$	$74.75\pm 0.50$	90.97 $\pm$ 14.65
MLP	$78.80\pm 0.92$	$82.63\pm 0.83$	$78.80\pm 0.92$	76.38 $\pm$ 0.99	$112.65\pm 16.28$
XGBoost	$77.79\pm 0.80$	$82.68\pm 0.63$	$77.79\pm 0.80$	$74.37\pm 0.72$	$91.60\pm 14.57$
LightGBM	$75.34\pm 0.78$	$79.94\pm 0.33$	$75.34\pm 0.78$	$72.07\pm 0.66$	$96.04\pm 14.61$
CatBoost	79.24 $\pm$ 0.26	83.82 $\pm$ 0.19	79.24 $\pm$ 0.26	$76.13\pm 0.33$	$99.94\pm 14.61$

TABLE II: Classification performance and total execution time on NSL-KDD across 30 pipeline-generation runs

Framework / Method	UNSW-NB15 (XGBoost)		NSL-KDD (MLP)
Framework / Method	Weighted $F_{1}$ (%)	Time (s)	Weighted $F_{1}$ (%)	Time (s)
LiteShield-MI-FS † [ 19 ]	78.17	114.07	67.34	96.47
XGBoost-FS † [ 20 ]	77.92	10.74	71.53	78.28
IGRF-RFE † [ 21 ]	78.13	857.20	–	–
DSSTE † [ 22 ]	–	–	73.59	77.56
AutoDP-LLM (Ours)	78.41 $\pm$ 0.22	106.38 $\pm$ 9.77	76.38 $\pm$ 0.99	$112.65\pm 16.28$

TABLE III: Performance and execution time comparison against other methods under the selected downstream classifiers: XGBoost for UNSW-NB15 and MLP for NSL-KDD

Ablation Variant	Component Configuration			Results (Weighted $F_{1}$ )
Ablation Variant	Semantic Guard	Semantic Pruning	RF-MI Completion	UNSW-NB15	NSL-KDD
w/o RF-MI Completion	$\checkmark$	$\checkmark$		77.28 $\pm$ 0.17	75.03 $\pm$ 0.47
w/o Semantic Pruning	$\checkmark$		$\checkmark$	77.90 $\pm$ 0.10	76.26 $\pm$ 1.14
w/o Semantic Components			$\checkmark$	78.33 $\pm$ 0.00	74.53 $\pm$ 0.00
Full method	$\checkmark$	$\checkmark$	$\checkmark$	78.41 $\pm$ 0.22	76.38 $\pm$ 0.99
XGBoost and MLP are used as the fixed representative classifiers for UNSW-NB15 and NSL-KDD, respectively.

TABLE IV: Component configurations and weighted

F_{1}

-scores of the feature-reduction ablation study

LLM Engine	Classifier	Classification Performance
LLM Engine	Classifier	Accuracy	Precision	Recall	$F_{1}$ -score
Claude-Sonnet-4.6	Random Forest	74.96 $\pm$ 0.31	83.38 $\pm$ 0.38	74.96 $\pm$ 0.31	77.74 $\pm$ 0.15
	XGBoost	74.84 $\pm$ 0.25	83.20 $\pm$ 0.48	74.84 $\pm$ 0.25	77.44 $\pm$ 0.18
	MLP	76.19 $\pm$ 0.95	81.47 $\pm$ 0.48	76.19 $\pm$ 0.95	77.13 $\pm$ 0.48
DeepSeek-V4-Pro	Random Forest	76.84 $\pm$ 0.17	82.31 $\pm$ 0.08	76.84 $\pm$ 0.17	78.56 $\pm$ 0.08
	XGBoost	77.23 $\pm$ 0.18	81.83 $\pm$ 0.13	77.23 $\pm$ 0.18	78.34 $\pm$ 0.11

TABLE V: Classification performance of the evaluated LLM engines on UNSW-NB15.

LLM Engine	Classifier	Classification Performance
LLM Engine	Classifier	Accuracy	Precision	Recall	$F_{1}$ -score
Claude-Sonnet-4.6	Random Forest	76.42 $\pm$ 0.05	82.33 $\pm$ 0.03	76.42 $\pm$ 0.05	72.81 $\pm$ 0.10
	XGBoost	78.67 $\pm$ 0.42	83.30 $\pm$ 0.22	78.67 $\pm$ 0.42	76.05 $\pm$ 0.41
	MLP	76.87 $\pm$ 0.26	79.73 $\pm$ 1.01	76.87 $\pm$ 0.26	74.82 $\pm$ 0.40
DeepSeek-V4-Pro	Random Forest	76.35 $\pm$ 0.63	82.13 $\pm$ 0.44	76.35 $\pm$ 0.63	72.41 $\pm$ 0.55
	XGBoost	77.31 $\pm$ 0.85	82.43 $\pm$ 0.33	77.31 $\pm$ 0.85	73.94 $\pm$ 1.00

TABLE VI: Classification performance of the evaluated LLM engines on NSL-KDD.

	Design-time Overhead			Run-time Efficiency
LLM Engine	Output Tokens	API Latency (s)	Cost (USD)	Number of Features	Training Time (s)
Dataset: UNSW-NB15 (Representative classifier: XGBoost)
Claude-Sonnet-4.6	18,199 $\pm$ 11,226	283.26 $\pm$ 99.65	$0.318$ \pm$ 0.168	$26\,\text{--}\,27$	14.73 $\pm$ 0.46
GPT-5.4	10,418 $\pm$ 2,139	60.23 $\pm$ 8.44	$0.194$ \pm$ 0.032	$26\,\text{--}\,28$	13.55 $\pm$ 0.45
DeepSeek-V4-Pro	79,152 $\pm$ 33,595	858.33 $\pm$ 354.50	$0.147$ \pm$ 0.059	$27\,\text{--}\,29$	14.19 $\pm$ 0.42
Dataset: NSL-KDD (Representative classifier: MLP)

TABLE VII: Operational efficiency of the evaluated LLM engines across datasets

Step / Agent	Parameter	Value	Design Rationale & Objective
Step 1: Analysis & Host Planning	Temperature	0.0	Promotes consistent planning and strict JSON schema compliance.
Step 2: Transformation & Resampling Specialist	Temperature	0.0	Eliminates syntax hallucinations and guarantees deterministic code execution.
Step 3: Feature-Reduction Specialist	Temperature	0.0	Ensures consistent feature selection and deterministic reduction logic.

TABLE VIII: Hyperparameter Configurations and Design Rationales for Multi-Agent Execution

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 11 — Fig. 3: Shared base context for all agents.

Panel of Figure 12 — Fig. 4: System prompt of Analysis Specialist Agent in Step 1.

Panel of Figure 13 — Fig. 5: System prompt of Transformation and Resampling Specialist Agent in Step 2.

Panel of Figure 14 — Fig. 6: System prompt of Feature Reduction Specialist Agent in Step 3.

Panel of Figure 15 — Fig. 7: Schema of the synthesized Validated Preprocessing Plan.

Explore similar work

GenTI: Benchmarking LLMs for Autonomous IDPS Rule Generation for Unseen Attacks
Jun 4, 2026Hassan Jalil Hadi, Rehana Yasmin, Ali ShokerIntrusion Detection Cyber Threat Intelligence
Cyber Security and Resilience Technology (CyberSaR), King Abdullah University of Science and Technology (KAUST), Saudi Arabia
Jev-IDS: System One Models for Network Intrusion Detection
Oct 1, 2026Paulo Severo, Silvio E. Quincozes, Amanda DiasIntrusion Detection Random Forest
Graduate Program in Software Engineering, Federal University of Pampa (UNIPAMPA), Alegrete, Rio Grande do Sul, Brazil
Evaluation of AutoML Frameworks for IDS under Imbalanced Data Conditions of the NSL-KDD Dataset
Jun 10, 2026Wiliane Carolina Silva, Evandro César Vilas Boas, Felipe A. P. de FigueiredoImbalanced Classification Intrusion Detection
Cybersecurity and Artificial Intelligence Laboratory (CS&I Lab) of the National Institute of Telecommunications (Inatel), Santa Rita do Sapucaí, Brazil · Wireless and Artificial Intelligence Laboratory (WAI Lab) of the National Institute of Telecommunications (Inatel), Santa Rita do Sapucaí, Brazil

Jun 4, 2026, cs.CRJ/K move · Enter open · S save

GenTI: Benchmarking LLMs for Autonomous IDPS Rule Generation for Unseen Attacks

Hassan Jalil Hadi, Rehana Yasmin, Ali Shoker

Cyber Security and Resilience Technology (CyberSaR), King Abdullah University of Science and Technology (KAUST), Saudi Arabia

Intrusion Detection Cyber Threat Intelligence

Open paper AskPDF arXiv

Rule-based Intrusion Detection and Prevention Systems (IDPS) offer precise attack detection as well as mitigation, however their manually crafted, signature-driven rules limit adaptability to emerging and zero-day threats. Additionally, existing public datasets (e.g., CICIDS2017, UNSW-NB15) focus on traffic classification and provide little structured information to support automatic rule synthesis or prevention logic. To address this gap, we propose Generative Thread Intelligence (GenTI) \footnote{GenTI refers to the proposed framework, and GTI refers to the dataset.} an LLM-driven benchmark for automatic generation of IDPS rules targeting unseen attacks. The dataset (GTI) aggregates over 150k detection and prevention rules from Snort, Suricata, Emerging Threats, as well as 50k YARA, each annotated with protocol behavior, payload signatures, contextual relationships, mappings to Cyber Threat Intelligence (CTI), along with actionable response types (alert, drop, reject). Moreover, on top of this corpus we design an LLM-based pipeline that transforms analyst prompts and representative payloads into deployable rules via structured prompt engineering, Chain-of-Thought (CoT) reasoning, as well as a Chain-of-Verification (CoVe) loop for syntactic, semantic, and security validation. The generated rules are executed in real time on (Snort/Suricata) and evaluated by syntax accuracy, semantic similarity, CTI coverage, security effectiveness as well as unseen attacks detection. Furthermore, our GenTI instantiation achieves a composite rule-quality score of 89.4%, with 94.8% CTI coverage, improving unseen attacks detection from 45% to 87.4% and reducing the false-positive rate from 8.5% to 2.3%. Overall, GenTI establishes the first large-scale benchmark that tightly couples rule-level CTI with LLM-based automation, enabling adaptive, self-evolving IDPS.

Abstract

Explore similar work

GenTI: Benchmarking LLMs for Autonomous IDPS Rule Generation for Unseen Attacks

Jev-IDS: System One Models for Network Intrusion Detection

Evaluation of AutoML Frameworks for IDS under Imbalanced Data Conditions of the NSL-KDD Dataset