arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2502.04643cs.LGcs.CLcs.CR
收藏

置信度引导:大语言模型的一种新攻击向量

Confidence Elicitation: A New Attack Vector for Large Language Models

National University of Singapore(新加坡国立大学);Institute of Data Science, National University of Singapore(新加坡国立大学数据科学研究所);Institute for Infocomm Research, A*STAR(新加坡科技研究局信息通信研究所);Centre for Frontier AI Research, A*STAR(新加坡科技研究局前沿人工智能研究中心)

Brian Formento, Chuan Sheng Foo, See-Kiong Ng

AI总结:

本文提出一种新的黑盒攻击范式,通过从大语言模型中引导出校准置信度作为攻击信号,最小化该置信度以提升错误分类概率,并在多个模型和数据集上取得优于现有硬标签黑盒攻击的效果。

摘要:

深度学习中的一个根本性问题是对抗鲁棒性。随着这些系统规模的扩大,此类问题依然存在。当前,拥有数十亿参数的大语言模型(LLMs)与它们早期较小的前身一样,同样遭受对抗攻击。然而,威胁模型已经发生了变化。此前,灰盒访问(即输入嵌入或输出logits/概率对用户可见)可能被认为是合理的。但随着闭源模型的引入,除了生成的输出之外,无法获得关于模型的任何信息。这意味着当前的黑盒攻击只能利用最终预测来判断攻击是否成功。在本工作中,我们研究并展示了在分类场景下仅拥有黑盒访问权限时,类似于利用输出概率进行攻击引导的潜力。这是通过从模型中引导出置信度的能力实现的。我们通过实验表明,对于当前的大语言模型,所引导出的置信度是经过校准的,并非幻觉。因此,通过最小化所引导出的置信度,我们可以增加错误分类的可能性。我们将所提出的技术与现有的引入词级替换的硬标签黑盒攻击方法进行比较,在两个模型(LLaMA-3-8B-Instruct和Mistral-7B-Instruct-V0.3)上的三个数据集上展示了具有前景的最新最优结果。

英文摘要:

A fundamental issue in deep learning has been adversarial robustness. As these systems have scaled, such issues have persisted. Currently, large language models (LLMs) with billions of parameters suffer from adversarial attacks just like their earlier, smaller counterparts. However, the threat models have changed. Previously, having gray-box access, where input embeddings or output logits/probabilities were visible to the user, might have been reasonable. However, with the introduction of closed-source models, no information about the model is available apart from the generated output. This means that current black-box attacks can only utilize the final prediction to detect if an attack is successful. In this work, we investigate and demonstrate the potential of attack guidance, akin to using output probabilities, while having only black-box access in a classification setting. This is achieved through the ability to elicit confidence from the model. We empirically show that the elicited confidence is calibrated and not hallucinated for current LLMs. By minimizing the elicited confidence, we can therefore increase the likelihood of misclassification. Our new proposed paradigm demonstrates promising state-of-the-art results on three datasets across two models (LLaMA-3-8B-Instruct and Mistral-7B-Instruct-V0.3) when comparing our technique to existing hard-label black-box attack methods that introduce word-level substitutions.

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。

↑