Abstract
This paper provides a comprehensive review of poisoning attacks against large language models (LLMs), drawing primarily from Fendley et al. (2025) and complementary studies from 2022–2025. It categorizes poisoning research into two key dimensions, Metrics and Specifications, to evaluate how attack success is measured and how attacks are implemented. This paper synthesizes quantitative results, experimental findings, and defense strategies across data, model, and multi-modal poisoning contexts. Finally, it highlights emerging challenges posed by self-adaptive and synthetic-data-driven LLMs, and proposes future research directions to strengthen model security and reliability.
Faculty Advisor/Mentor
Yan Lu
Document Type
Paper
Disciplines
Cybersecurity
DOI
10.25776/7avg-dc64
Publication Date
11-10-2025
Upload File
wf_yes
Included in
A Systematic Review of Poisoning Attacks Against Large Language Models (LLM)
This paper provides a comprehensive review of poisoning attacks against large language models (LLMs), drawing primarily from Fendley et al. (2025) and complementary studies from 2022–2025. It categorizes poisoning research into two key dimensions, Metrics and Specifications, to evaluate how attack success is measured and how attacks are implemented. This paper synthesizes quantitative results, experimental findings, and defense strategies across data, model, and multi-modal poisoning contexts. Finally, it highlights emerging challenges posed by self-adaptive and synthetic-data-driven LLMs, and proposes future research directions to strengthen model security and reliability.