Is poisoning a real threat to LLM alignment? Maybe more so than you think

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Pathmanathan, Pankayaraj, Chakraborty, Souradip, Liu, Xiangyu, Liang, Yongyuan, Huang, Furong
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866916783188869120
author Pathmanathan, Pankayaraj
Chakraborty, Souradip
Liu, Xiangyu
Liang, Yongyuan
Huang, Furong
author_facet Pathmanathan, Pankayaraj
Chakraborty, Souradip
Liu, Xiangyu
Liang, Yongyuan
Huang, Furong
contents Recent advancements in Reinforcement Learning with Human Feedback (RLHF) have significantly impacted the alignment of Large Language Models (LLMs). The sensitivity of reinforcement learning algorithms such as Proximal Policy Optimization (PPO) has led to new line work on Direct Policy Optimization (DPO), which treats RLHF in a supervised learning framework. The increased practical use of these RLHF methods warrants an analysis of their vulnerabilities. In this work, we investigate the vulnerabilities of DPO to poisoning attacks under different scenarios and compare the effectiveness of preference poisoning, a first of its kind. We comprehensively analyze DPO's vulnerabilities under different types of attacks, i.e., backdoor and non-backdoor attacks, and different poisoning methods across a wide array of language models, i.e., LLama 7B, Mistral 7B, and Gemma 7B. We find that unlike PPO-based methods, which, when it comes to backdoor attacks, require at least 4\% of the data to be poisoned to elicit harmful behavior, we exploit the true vulnerabilities of DPO more simply so we can poison the model with only as much as 0.5\% of the data. We further investigate the potential reasons behind the vulnerability and how well this vulnerability translates into backdoor vs non-backdoor attacks.
format Preprint
id arxiv_https___arxiv_org_abs_2406_12091
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Is poisoning a real threat to LLM alignment? Maybe more so than you think
Pathmanathan, Pankayaraj
Chakraborty, Souradip
Liu, Xiangyu
Liang, Yongyuan
Huang, Furong
Machine Learning
Computation and Language
Cryptography and Security
Recent advancements in Reinforcement Learning with Human Feedback (RLHF) have significantly impacted the alignment of Large Language Models (LLMs). The sensitivity of reinforcement learning algorithms such as Proximal Policy Optimization (PPO) has led to new line work on Direct Policy Optimization (DPO), which treats RLHF in a supervised learning framework. The increased practical use of these RLHF methods warrants an analysis of their vulnerabilities. In this work, we investigate the vulnerabilities of DPO to poisoning attacks under different scenarios and compare the effectiveness of preference poisoning, a first of its kind. We comprehensively analyze DPO's vulnerabilities under different types of attacks, i.e., backdoor and non-backdoor attacks, and different poisoning methods across a wide array of language models, i.e., LLama 7B, Mistral 7B, and Gemma 7B. We find that unlike PPO-based methods, which, when it comes to backdoor attacks, require at least 4\% of the data to be poisoned to elicit harmful behavior, we exploit the true vulnerabilities of DPO more simply so we can poison the model with only as much as 0.5\% of the data. We further investigate the potential reasons behind the vulnerability and how well this vulnerability translates into backdoor vs non-backdoor attacks.
title Is poisoning a real threat to LLM alignment? Maybe more so than you think
topic Machine Learning
Computation and Language
Cryptography and Security
url https://arxiv.org/abs/2406.12091