Infusion: 通过影响函数编辑训练数据来塑造模型行为
Infusion: Shaping Model Behavior by Editing Training Data via Influence Functions
AI总结:
Infusion通过影响函数近似计算训练数据的微小扰动,以诱导模型行为的针对性变化。在CIFAR-10上,仅对0.2%的训练数据进行细微编辑即可与基线方法竞争,且在不同架构间具有迁移性。
摘要:
影响函数常用于将模型行为归因于训练文档。我们探索其反向应用:通过编辑训练数据来诱导模型行为。我们的框架Infusion利用可扩展的影响函数近似计算训练文档的微小扰动,通过参数偏移诱导模型行为的针对性变化。我们在视觉和语言领域跨数据中毒任务上评估了Infusion。在CIFAR-10上,我们展示了通过Infusion对仅0.2%(100/45,000)的训练文档进行细微编辑,可以与基线方法(插入少量显式行为示例)竞争。我们还发现Infusion在不同架构(ResNet ↔ CNN)间具有迁移性,表明单一中毒语料库可以影响多个独立训练的模型。在初步语言实验中,我们确定了我们的方法增加目标行为概率和失败的情况,发现它最有效于放大模型已学习的行为。总体而言,这些结果表明,对训练数据进行微小、细微的编辑可以系统地塑造模型行为,凸显了训练数据可解释性对对手和防御者的重要性。我们提供代码:https://github.com/jrosseruk/infusion.
英文摘要:
Influence functions are commonly used to attribute model behavior to training documents. We explore the reverse: crafting training data that induces model behavior. Our framework, Infusion, uses scalable influence-function approximations to compute small perturbations to training documents that induce targeted changes in model behavior through parameter shifts. We evaluate Infusion on data poisoning tasks across vision and language domains. On CIFAR-10, we show that making subtle edits via Infusion to just 0.2% (100/45,000) of the training documents can be competitive with the baseline of inserting a small number of explicit behavior examples. We also find that Infusion transfers across architectures (ResNet $\leftrightarrow$ CNN), suggesting a single poisoned corpus can affect multiple independently trained models. In preliminary language experiments, we characterize when our approach increases the probability of target behaviors and when it fails, finding it most effective at amplifying behaviors the model has already learned. Taken together, these results show that small, subtle edits to training data can systematically shape model behavior, underscoring the importance of training data interpretability for adversaries and defenders alike. We provide the code here: https://github.com/jrosseruk/infusion.
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。