arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2402.01920cs.LGcs.AIcs.CL
收藏

奖励模型学习中的偏好投毒攻击

Preference Poisoning Attacks on Reward Model Learning

Washington University in St. Louis(圣路易斯华盛顿大学);University of Wisconsin-Madison(威斯康星大学麦迪逊分校)

Junlin Wu, Jiongxiao Wang, Chaowei Xiao, Chenguang Wang, Ning Zhang, Yevgeniy Vorobeychik

AI总结:

该研究针对奖励模型偏好学习的投毒漏洞,提出基于梯度和按距离排序的两类攻击方法,实验证明少量投毒即可实现高攻击成功率,且现有前沿防御方法效果有限。

摘要:

从成对比较中学习奖励模型是自主控制、对话智能体、推荐系统等多个领域的核心组件,服务于让自动化决策与用户偏好对齐的广义目标。这类方法需要收集人类偏好信息,反馈通常以匿名方式提供。由于偏好具有主观性,不存在可用于比对的金标准;但高影响力系统对偏好学习的依赖,促使恶意行为者有强烈动机篡改以此方式收集的数据,以达成自身目的。我们通过设定一名可翻转少量偏好比较样本、从而提升或降低目标结果排名的攻击者,研究了此漏洞的性质与影响范围。我们提出两类用于此类攻击的算法方案:基于梯度的框架,以及多种按距离排序(rank-by-distance)方法的变体。随后,我们在来自三个领域的数据集上评估了两类方法中最优攻击实现恶意目标的效果,这三个领域分别是自主控制、推荐系统以及文本提示-回复偏好学习。我们发现最优攻击通常成功率极高,最极端的情况下仅需投毒0.3%的数据就能达到100%的成功率。但哪种攻击效果最佳在不同领域间差异显著。此外,我们观察到更简单、可扩展性更强的按距离排序方法通常能与基于梯度的方法效果相当,有时甚至显著优于后者。最后,我们证明针对其他类别投毒攻击的前沿防御方法在我们的场景中效果有限。

英文摘要:

Learning reward models from pairwise comparisons is a fundamental component in a number of domains, including autonomous control, conversational agents, and recommendation systems, as part of a broad goal of aligning automated decisions with user preferences. These approaches entail collecting preference information from people, with feedback often provided anonymously. Since preferences are subjective, there is no gold standard to compare against; yet, reliance of high-impact systems on preference learning creates a strong motivation for malicious actors to skew data collected in this fashion to their ends. We investigate the nature and extent of this vulnerability by considering an attacker who can flip a small subset of preference comparisons to either promote or demote a target outcome. We propose two classes of algorithmic approaches for these attacks: a gradient-based framework, and several variants of rank-by-distance methods. Next, we evaluate the efficacy of best attacks in both these classes in successfully achieving malicious goals on datasets from three domains: autonomous control, recommendation system, and textual prompt-response preference learning. We find that the best attacks are often highly successful, achieving in the most extreme case 100\% success rate with only 0.3\% of the data poisoned. However, \emph{which} attack is best can vary significantly across domains. In addition, we observe that the simpler and more scalable rank-by-distance approaches are often competitive with, and on occasion significantly outperform, gradient-based methods. Finally, we show that state-of-the-art defenses against other classes of poisoning attacks exhibit limited efficacy in our setting.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑