RAG Poisoning Attack (Document-Level Targeting) on FinQA
47.1RSR@5CorpusPoisoning
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| CorpusPoisoningRetriever=Qwen3-Embedding-8B, Backend LLM=GPT-4o mini2025.12 | 47.1 | 21.02 | 25.9 | 3.11 | |
| Prompt InjectionRetriever=Qwen3-Embedding-8B, Backend LLM=GPT-4o mini2025.12 | 43.7 | 22.2 | 36.84 | 4.98 | |
| PoisonedRAG-BRetriever=Qwen3-Embedding-8B, Backend LLM=GPT-4o mini2025.12 | 42.9 | 28.67 | 39.16 | 5.35 | |
| MIRAGERetriever=Qwen3-Embedding-8B, Backend LLM=GPT-4o mini2025.12 | 40.5 | 41.48 | 42.73 | 5.83 | |
| PARADOXRetriever=Qwen3-Embedding-8B, Backend LLM=GPT-4o mini2025.12 | 6.6 | 4.55 | 21.21 | 5.27 | |
| DIGARetriever=Qwen3-Embedding-8B, Backend LLM=GPT-4o mini2025.12 | 2.9 | 0 | 0 | 1.8 | |
| GCG AttackRetriever=Qwen3-Embedding-8B, Backend LLM=GPT-4o mini2025.12 | 0 | — | — | 1.66 |