Best-of-Venom: Attacking RLHF by Injecting Poisoned Preference Data

arXiv:2404.05530v1 Announce Type: cross
Abstract: Reinforcement Learning from Human Feedback (RLHF) is a popular method for aligning Language Models (LM) with human values and preferences. RLHF requires a large number of preference pairs as training data, which are often used in both the Supervised Fine-Tuning and Reward Model training, and therefore publicly available datasets are commonly used. In this work, we study to what extent a malicious actor can manipulate the LMs generations by poisoning the preferences, i.e., injecting poisonous …

abstract arxiv cs.ai cs.cl cs.cr cs.lg data feedback fine-tuning human human feedback language language models popular reinforcement reinforcement learning reward model rlhf supervised fine-tuning training training data type values