MEL Adversarial Example Dataset
收藏资源简介:
本数据集是首个针对多模态实体链接(MEL)模型对抗鲁棒性的评估数据集,旨在解决现有MEL模型在视觉对抗攻击下的鲁棒性不足问题。该数据集基于五个常见MEL基准数据集构建,包括Wikidata-MEL、Richpedia-MEL、WikiDiverse、WIKIPerson和M3EL,覆盖了多种对抗攻击场景,如图像到文本(I2T)和图像+文本到文本(IT2T)等任务。数据集通过引入对抗样本,评估了现有MEL模型的鲁棒性,并为未来加强多模态系统在对抗环境中的鲁棒性研究奠定了基础。
This dataset is the first evaluation dataset for the adversarial robustness of multimodal entity linking (MEL) models, aiming to address the insufficient robustness of existing MEL models under visual adversarial attacks. Constructed based on five mainstream MEL benchmark datasets including Wikidata-MEL, Richpedia-MEL, WikiDiverse, WIKIPerson and M3EL, it covers various adversarial attack scenarios and tasks such as image-to-text (I2T) and image+text-to-text (IT2T). The dataset evaluates the robustness of existing MEL models by introducing adversarial samples, and lays a solid foundation for future research on enhancing the adversarial robustness of multimodal systems in adversarial environments.

- 1通过北京大学, 杭州智源研究院有限公司, 中国机械科学研究总院, 北京大数据先进研究院 · 2025年



