"PromptGame Dataset: Prompt Injection Attack Benchmark and Defense Evaluation Data"
收藏DataCite Commons2026-02-15 更新2026-05-03 收录
下载链接：
https://ieee-dataport.org/documents/promptgame-dataset-prompt-injection-attack-benchmark-and-defense-evaluation-data
下载链接
链接失效反馈官方服务：
资源简介：
"This dataset accompanies the paper \"Semantic Integrity Under Prompt Injection: A Game-Theoretic Analysis of Equilibria in Adversarial Natural Language Processing Systems\" published in IEEE Transactions on Information Forensics and Security.The dataset contains: (1) A comprehensive benchmark of 50 prompt injection attack types across four categories\u2014direct injection (15 types from TensorTrust), RAG poisoning (12 types from PoisonedRAG), separator\/delimiter exploits (13 types from Li et al.), and cascading agent attacks (10 types from InjecAgent)\u2014with 150 concrete attack instances; (2) Evaluation results from experiments across four LLM architectures (GPT-4o, Claude 3, LLaMA-3-70B, Mistral-7B) testing three defense mechanisms (Semantic Prompt Binding, Adaptive Risk Assessment, Randomized Token Embedding Shuffling); (3) Metrics including Attack Success Rate, Semantic Fidelity, defense overhead, and equilibrium satisfaction rates (SRE, ISE, CRE); (4) System prompts used for evaluation across 10 applic"
提供机构：
IEEE DataPort创建时间：
2026-02-15
