【IEEE Access2021】Adversarial Attacks Against Face Recognition_A_Comprehensive_Study
1. 相关工作
- 介绍了目前常用的人脸识别的models,包括Facenet,VGG-face,Cosface,Arcface等等,后两者是基于相似度cosine.
- Dataset:LFW,CASIA-Webface,MS-Celeb-1M,VGGface2,Megaface。
2. 对抗攻击方法
FGSM,L-BFGS,I-FGSM, JSMA,One-pixel attack, DeepFool,C&W等
人脸对抗攻击
- Evolutionary attack (1+1-CMS-ES)降维搜索【Efficient Decision-based Black-box Adversarial Attacks on Face Recognition】
- FIM:
- 对抗眼镜(1,2)【Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition】【A General Framework for Adversarial Examples with Objectives】
- VLA-light-based:【VLA: A Practical Visible Light-based Attack on Face Recognition Systems in Physical World】
-
P-FGVM(惩罚FGVM):就是在FGVM基础上加一个lamda*(X-X^)更新loss
【Adversarial face de-identification】
-
FLM(Face landmark manipulation),基于几何的方法,通过面部特征点的移动,最后提出一个Semantic group用于分组更新【Fast Geometrically-Perturbed Adversarial Faces】
-
DFAnet,在model中加入dropout,用于随机丢失部分特征,来提高transferability【Towards Transferable Adversarial Attack against Deep Face Recognition】
3. 对抗防御方法
基本上是三种类型——调整训练/测试阶段的输入、调整model网络结构,使用外部model