A SMALL LANGUAGE MODEL THAT SPECIALIZES IN CYBERATTACKS WITH AUTONOMOUS FINE-TUNING SUPERVISION
DOI:
https://doi.org/10.66104/yv01hn69Keywords:
Small Language Model, fine-tuning, CybersecurityAbstract
Alert overload in Security Operations Centers and privacy concerns in cloud-based artificial intelligence solutions are relevant challenges for log analytics in enterprise environments. This work proposes a machine learning architecture for semantic screening of network logs, based on the fine-tuning of a small language model (SLM) executed locally. The methodology adopts the LLM judge paradigm, in which a language model is used both to generate synthetic training data and to automatically evaluate SLM responses. This strategy reduces reliance on manual labeling, preserves data confidentiality, and enables straightforward processing at the network edge. The experimental evaluation, carried out through an automated pipeline, considered criteria of logical accuracy and structural fidelity. The results indicate that the adjusted model achieved a flow rate of 103 Tokens per Second (TPS) and 100% metric fidelity, surpassing generic models in accuracy and efficiency for security applications in local networks.
Downloads
References
BANDARU, R. Cloud-native microservices with Docker and Kubernetes: build and deploy scalable microservices using Docker, Kubernetes, and Helm. 2. ed. Birmingham: Packt Publishing, 2022.
ABUSEIPDB. AbuseIPDB: IP address reputation tracking. Disponível em: https://www.abuseipdb.com/. Acessado em: Abril de 2026.
ARXIV. Knowledge-to-Data: LLM-Driven Synthesis of Structured Network Traffic for Testbed-Free IDS Evaluation. arXiv preprint arXiv:2601.05022, 2026.
BANG, Y. et al. CacheGen: Fast Context Loading for Language Model Applications. arXiv preprint arXiv:2310.07240, 2023.
CAI, Z. et al. Graph-based network anomaly detection: A survey. Computer Networks, 2021.
DAMIR, A. et al. Debiasing LLM-as-a-Judge in Cybersecurity: A Study of Judge Reliability and Human Alignment. In: OpenReview, 2025.
DETTMERS, T. et al. QLoRA: Efficient Finetuning of Quantized LLMs. In: NeurIPS, 2023. DOI: https://doi.org/10.52202/075280-0441
ES, S. et al. Ragas: Automated Evaluation of Retrieval Augmented Generation. arXiv preprint arXiv:2309.15217, 2023. DOI: https://doi.org/10.18653/v1/2024.eacl-demo.16
FOUNDERTOFORTUNE. Contract-based AI: Moving beyond Prompt Engineering for Reliable Systems. AI Software Engineering Review, 2025.
JOHNSON, J. et al. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 2019.
KARRAS, A. et al. LLMs for Cybersecurity in the Big Data Era: A Comprehensive Review. MDPI, 2025. DOI: https://doi.org/10.3390/info16110957
LEVI, M. et al. Toward Cybersecurity-Expert Small Language Models. arXiv preprint arXiv:2510.14113, 2025.
LEWIS, P. et al. Retrieval-augmented generation for knowledge-intensive NLP tasks. In: NeurIPS, 2020.
MA, L. et al. AdaptiveLog: An Adaptive Log Analysis Framework with the Collaboration of Large and Small Language Model. arXiv preprint arXiv:2501.11031, 2025.
MASRI, Y. et al. Benchmarking Small Language Models and Small Reasoning Language Models on System Log Severity Classification. arXiv preprint arXiv:2601.07790, 2026.
MDPI. Domain Knowledge-Infused Synthetic Data Generation for LLM-Based ICS Intrusion Detection: Mitigating Data Scarcity and Imbalance. Electronics, v. 15, 2026. DOI: https://doi.org/10.3390/electronics15020371
MUZAMMIL, S. et al. Towards Small Language Models for Security Query Generation in SOC Workflows. arXiv preprint arXiv:2512.06660, 2025.
PALMA, G. et al. Leveraging Large Language Models for Scalable and Explainable Cybersecurity Log Analysis. MDPI, 2024. DOI: https://doi.org/10.3390/jcp5030055
PALO ALTO NETWORKS RESEARCH. Why SLMs are Replacing LLMs for Real-time Log Analysis in 2025. Cyber Defense Insights, 2025.
PRIETO, P. et al. Edge Deployment of Small Language Models, a comprehensive comparison of CPU, GPU and NPU backends. arXiv preprint arXiv:2511.22334, 2025.
REINSEL, D. et al. Data Age 2025: The Digitization of the World from Edge to Core. IDC, 2018.
RODRIGUES, R. F. et al. Análise de logs de rede sem fio com suporte à tomada de decisão na infraestrutura Tecnologia da informação. Revista Brasileira de Criminalística, v. 14, 2025. DOI: https://doi.org/10.15260/rbc.v14i1.893
SAMPAIO, I. P. Automação de Processos de Resposta a Eventos de Segurança. 2024.
STROM, B. E. et al. Mitre att&ck: Design and philosophy. The MITRE Corporation, 2018.
TALASSO, G. U. et al. Fine-Tuning Eficiente de Modelos de Linguagem Para Detectar Anomalias em Logs Privados usando Aprendizado Federado. In: SIMPÓSIO BRASILEIRO DE REDES DE COMPUTADORES E SISTEMAS DISTRIBUÍDOS (SBRC), 2025. DOI: https://doi.org/10.5753/sbrc.2025.5859
TANIUM RESEARCH. The Intent Gap: Challenges in Autonomous AI Security Agents. Global Cybersecurity Trends, 2026.
TARIQ, M. et al. Large Language Models in Cybersecurity: A Survey of Applications, Vulnerabilities, and Defense Techniques. MDPI, 2025.
WANG, Y. et al. Self-Instruct: Aligning Language Models with Self-Generated Instructions. In: ACL, 2023. DOI: https://doi.org/10.18653/v1/2023.acl-long.754
WEI, B. et al. CORTEX: Collaborative LLM Agents for High-Stakes Alert Triage. arXiv preprint arXiv:2510.00311, 2025.
WEI, J. et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In: NeurIPS, 2022.
WIZ RESEARCH. Fine-Tuning a Small Language Model for Secrets Detection in Code. Wiz, 2025.
WU, J. et al. When Cache Poisoning Meets LLM Systems: Semantic Cache Poisoning. In: NDSS Symposium, 2026. DOI: https://doi.org/10.14722/ndss.2026.240200
YANG, L. et al. True Attacks, Attack Attempts, or Benign Triggers? Empirical Measurement of Network Alerts in a Security Operations Center. In: 33rd USENIX Security Symposium, 2024.
ZHANG, J. et al. When LLMs meet cybersecurity: a systematic literature review. Cybersecurity, 2025. DOI: https://doi.org/10.1186/s42400-025-00361-w
ZHENG, L. et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. In: Neural Information Processing Systems (NeurIPS), 2023.
ZHONG, A. et al. LogParser-LLM: Advancing Efficient Log Parsing with Large Language Models. arXiv preprint, 2024. DOI: https://doi.org/10.1145/3637528.3671810
ZHOU, J. et al. Instruction-Following Evaluation for Large Language Models. arXiv preprint arXiv:2311.07911, 2023.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Gustavo Cavalcante Correa, João Gabriel de Oliveira Silva, Nicollas Aguiar Avila, Allan Costa, Marcos Melo, Matheus Brito, Hugo Leonardo Melo dos Santos

This work is licensed under a Creative Commons Attribution 4.0 International License.
Authors who publish in this journal agree to the following terms:
Authors retain copyright and grant the journal the right of first publication, with the work simultaneously licensed under the Creative Commons Attribution License, which permits the sharing of the work with proper acknowledgment of authorship and initial publication in this journal;
Authors are authorized to enter into separate, additional agreements for the non-exclusive distribution of the version of the work published in this journal (e.g., posting in an institutional repository or publishing it as a book chapter), provided that authorship and initial publication in this journal are properly acknowledged, and that the work is adapted to the template of the respective repository;
Authors are permitted and encouraged to post and distribute their work online (e.g., in institutional repositories or on their personal websites) at any point before or during the editorial process, as this may lead to productive exchanges and increase the impact and citation of the published work (see The Effect of Open Access);
Authors are responsible for correctly providing their personal information, including name, keywords, abstracts, and other relevant data, thereby defining how they wish to be cited. The journal’s editorial board is not responsible for any errors or inconsistencies in these records.
PRIVACY POLICY
The names and email addresses provided to this journal will be used exclusively for the purposes of this publication and will not be made available for any other purpose or to third parties.
Note: All content of the work is the sole responsibility of the author and the advisor.
