Skip to content

Black-Box Adversarial Attack Competition Based on CIFAR-10 Dataset.

Notifications You must be signed in to change notification settings

ghy-cmd/Intelligent-Security-Competition

Repository files navigation

Intelligent-Security-Competition

🐟背景

image-20240116125537685

通过引入对抗样本,攻击者很容易就可以通过肉眼几乎观察不到的微小扰动,使模型分类失误。

🚀主要贡献

image-20240116125953695

  • 防御模型的集成可以在一定程度上提高整体的鲁棒性,我们决定同时集成来自ARES库的6个防御模型和RoubustBench排行榜上的4个防御模型
  • 攻击方法选择了包含L2PGD攻击方法在内的组合攻击方法AutoAttack,用于攻击模型以生成对抗样本
  • 二分空间中,最优参数搜索和参数动态适应优化

🚲实验结果

在《 Cifar-10 数字世界 无限制对抗攻击竞赛》中排名TOP1

image-20240116130509391

❗声明

相关模型权重和关键代码未上传,后续课程禁止复用抄袭本仓库代码

About

Black-Box Adversarial Attack Competition Based on CIFAR-10 Dataset.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published