EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A Multi-Perspective Benchmark Dataset and Moderation Model for LLM Safety Evaluation with Adversarial Robustness Analysis

Naseem Machlovi, Maryam Saleki, Ruhul Amin, Mohamed Rahouti, Shawqi Al-Maliki, Junaid Qadir, Mohamed Abdallah, Ala Al-Fuqaha

ACM Transactions on Social Computing · 30.06.2026 · 9 (2) · 1-32

DOI: 10.1145/3815159 ↗

Kurz gefasst

Fachartikel in der Zeitschrift ACM Transactions on Social Computing, online erschienen am 30.05.2026. Verfasst von Naseem Machlovi und 7 weiteren Autor:innen. Die Arbeit nennt 50 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Qatar Research Development and Innovation Council.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

As large language models (LLMs) become deeply embedded in daily life, the urgent need for safer moderation systems that distinguish between naive and harmful requests while upholding appropriate censorship boundaries has never been greater. While existing LLMs can detect dangerous or unsafe content, they often struggle with nuanced cases such as implicit offensiveness, subtle gender and racial biases, and jailbreak prompts, due to the subjective and context-dependent nature of these issues. Furthermore, their heavy reliance on training data can reinforce societal biases, resulting in inconsistent and ethically problematic outputs. To address these challenges, we introduce GuardEval, a unified multi-perspective benchmark dataset designed for both training and evaluation, containing 106 fine-grained categories spanning human emotions, offensive and hateful language, gender and racial bias, and broader safety concerns. We also present GemmaGuard (GGuard), a Quantized Low-Rank Adaptation (QLoRA), fine-tuned version of Gemma3-12B trained on GuardEval, to assess content moderation with fine-grained labels. Our evaluation shows that GGuard achieves a macro F1 score of 0.832, substantially outperforming leading moderation models, including OpenAI Moderator (0.64) and Llama Guard (0.61). We show that multi-perspective, human-centered safety benchmarks are critical for mitigating inconsistent moderation decisions. GuardEval and GGuard together demonstrate that diverse, representative data materially improve safety, and adversarial robustness on complex, borderline cases.

Förderung

  • Qatar Research Development and Innovation Council

Zitieren

APA

Machlovi, N., Saleki, M., Amin, R., Rahouti, M., Al-Maliki, S., Qadir, J., Abdallah, M., & Al-Fuqaha, A. (2026). A Multi-Perspective Benchmark Dataset and Moderation Model for LLM Safety Evaluation with Adversarial Robustness Analysis. ACM Transactions on Social Computing, 9(2), 1-32. https://doi.org/10.1145/3815159

Vancouver

Machlovi N, Saleki M, Amin R, Rahouti M, Al-Maliki S, Qadir J, et al. A Multi-Perspective Benchmark Dataset and Moderation Model for LLM Safety Evaluation with Adversarial Robustness Analysis. ACM Transactions on Social Computing. 2026;9(2):1-32. doi:10.1145/3815159

BibTeX
@article{Machlovi2026,
  author = {Machlovi, Naseem and Saleki, Maryam and Amin, Ruhul and Rahouti, Mohamed and Al-Maliki, Shawqi and Qadir, Junaid and Abdallah, Mohamed and Al-Fuqaha, Ala},
  title = {A Multi-Perspective Benchmark Dataset and Moderation Model for LLM Safety Evaluation with Adversarial Robustness Analysis},
  journal = {ACM Transactions on Social Computing},
  year = {2026},
  volume = {9},
  number = {2},
  pages = {1--32},
  doi = {10.1145/3815159}
}
Kostenlos starten →