Skip to content
InfoResearchPeer-reviewedLLM-specific

How Dangerous Are Data Poisoning Attacks on LLMs? An Empirical Study

Published
Record updated
View JSON

Summary

This empirical study tests trigger-based data poisoning on LLaMA-3 8B Instruct, using 30,220 training examples from ten SuperNaturalInstructions tasks. Under a corrected attribution procedure, a 2% removal budget identified 61 of 604 poisoned examples, giving 10.10% precision and recall, while a literal search using the known trigger found all 604 with no false positives. The poisoned adapter produced the exact target response on 94.0–100.0% of triggered non-target examples across datasets.