上亿级神经网络对抗防御数据集（查询请求）

本数据集面向大规模深度神经网络的对抗鲁棒性研究构建。深度神经网络在多个领域展现出卓越性能，但其存在的对抗样本脆弱性严重威胁模型安全性。本研究针对黑盒查询攻击场景，提出一种简单有效的对抗防御算法，并在视觉Transformer架构ViT-L-16上进行验证。实验结果表明，该防御机制可显著提升模型鲁棒性，与无防御基准相比，攻击成功所需的平均查询次数显著增加逾8000次（Square Attack基准测试）。为推进对抗防御算法的研究，项目组构建了亿级规模的对抗防御基准数据集，主要包括算法代码、模型权重等数据，为大模型对抗鲁棒性研究提供了基准方法与技术基础。This dataset is constructed for research on adversarial robustness of large-scale deep neural networks. Deep neural networks have exhibited outstanding performance across multiple domains, but their vulnerability to adversarial examples poses a serious threat to model security. This study proposes a simple and effective adversarial defense algorithm targeting black-box query attack scenarios, and validates it on the Vision Transformer architecture ViT-L-16. Experimental results show that this defense mechanism can significantly enhance model robustness: compared with the undefended baseline, the average number of queries required for a successful attack increases by more than 8000 under the Square Attack benchmark test. To promote research on adversarial defense algorithms, the project team has constructed a hundred-million-scale adversarial defense benchmark dataset, which mainly includes data such as algorithm codes and model weights, providing benchmark methods and technical foundations for research on adversarial robustness of large language models (LLMs).

搜集汇总

背景与挑战

背景概述

该数据集面向大规模深度神经网络的对抗鲁棒性研究，针对黑盒查询攻击场景构建。它包含一种经实验验证可显著提升模型鲁棒性的防御算法及相关基准数据，数据规模达亿级，为相关研究提供了方法与技术基础。