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 2510.07192cs.LG
收藏

对大型语言模型的投毒攻击需要近乎恒定的投毒样本数量

Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples

UK AI Security Institute(英国人工智能安全研究所);Anthropic;Alan Turing Institute(艾伦·图灵研究所);OATML, University of Oxford(OATML,牛津大学);ETH Zurich(苏黎世联邦理工学院)
Alexandra Souly, Javier Rando, Ed Chapman, Xander Davies, Burak Hasircioglu, Ezzeldin Shereen, Carlos Mougan, Vasilios Mavroudis, Erik Jones, Chris Hicks, Nicho… 展开作者

Alexandra Souly, Javier Rando, Ed Chapman, Xander Davies, Burak Hasircioglu, Ezzeldin Shereen, Carlos Mougan, Vasilios Mavroudis, Erik Jones, Chris Hicks, Nicholas Carlini, Yarin Gal, Robert Kirk

AI总结:

本研究首次证明,投毒攻击大型语言模型所需的投毒文档数量近乎恒定,不随模型或数据集规模增长,且250个文档即可在6亿至130亿参数模型上成功注入后门,凸显了防御研究的紧迫性。

摘要:

投毒攻击可以通过向大型语言模型(LLM)的训练数据中注入恶意文档来危害其安全性。现有工作研究了预训练投毒,假设攻击者控制训练语料库的一定百分比。然而,对于大型模型,即使很小的百分比也转化为不切实际的大量数据。这项工作首次证明,投毒攻击所需的文档数量与数据集大小无关,近乎恒定。我们进行了迄今为止最大规模的预训练投毒实验,在chinchilla最优数据集(60亿至2600亿个token)上预训练了参数从6亿到130亿的模型。我们发现,250个投毒文档能够同样地损害所有模型和数据集大小的模型,尽管最大的模型在超过20倍的干净数据上进行训练。我们还进行了较小规模的实验,以消融可能影响攻击成功的因素,包括更广泛的投毒与干净数据比例以及投毒样本的非随机分布。最后,我们证明了在微调期间投毒具有相同的动态特性。总的来说,我们的结果表明,通过数据投毒注入后门可能比以前认为的更容易,因为所需的投毒数量不随模型大小而扩展,这凸显了在未来模型中需要更多关于防御措施的研究以减轻这种风险。

英文摘要:

Poisoning attacks can compromise the safety of large language models (LLMs) by injecting malicious documents into their training data. Existing work has studied pretraining poisoning assuming adversaries control a percentage of the training corpus. However, for large models, even small percentages translate to impractically large amounts of data. This work demonstrates for the first time that poisoning attacks instead require a near-constant number of documents regardless of dataset size. We conduct the largest pretraining poisoning experiments to date, pretraining models from 600M to 13B parameters on chinchilla-optimal datasets (6B to 260B tokens). We find that 250 poisoned documents similarly compromise models across all model and dataset sizes, despite the largest models training on more than 20 times more clean data. We also run smaller-scale experiments to ablate factors that could influence attack success, including broader ratios of poisoned to clean data and non-random distributions of poisoned samples. Finally, we demonstrate the same dynamics for poisoning during fine-tuning. Altogether, our results suggest that injecting backdoors through data poisoning may be easier for large models than previously believed as the number of poisons required does not scale up with model size, highlighting the need for more research on defences to mitigate this risk in future models.

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。

↑