2021WSB-day5-3:Sébastien讲解Biometric Presentation Attack Detection
作者:互联网
来自瑞士IDIAP研究所研究员、欧盟FP7项目负责人Sébastien Marcel讲解生物特征识别的 Presentation Attack Detection。
文章目录
- outline
- 现实的例子
- 定义
- 一些attack
- 没关系啊,红外相机可以检测到
- 化妆易容术
- Magic passport 就是一个护照可以和两个人的照片合上
- deep fake
- 那么怎么检测呢?
- 我们也有数据集
- 我们的工作
- PAD methods
- 问
outline
Introduction
Presentation Attacks in reality
Definitions
Presentation Attacks (PAs)
Face PAs
Face PAD
Biometrics and PAD
References
现实的例子
定义
一些attack
假的指纹。
讲了很多attack,看你需要多少钱的:)
没关系啊,红外相机可以检测到
化妆易容术
化妆就有问题了!!
Magic passport 就是一个护照可以和两个人的照片合上
- 这个文章我读过啦。 这个只是假设啦,论文里面提出来的,不是真实的
deep fake
那么怎么检测呢?
- Eye-blinking
- Motion
- Texture analysis
- Frequency analysis
- Image Quality Analysis
- Make-up PAD with fusion of holistic and local CNNs
- Deep Pixel-wise
- Multi-Spectral PAD
我们也有数据集
我们的工作
两个模块:
We measure 2 errors:
False Match Rate (FMR): zero-effort impostors incorrectly
matched as genuines – also referred to as False Acceptance
Rate (FAR)
False Non-Match Rate (FNMR): genuines not matched – also
referred to as False Rejection Rate (FRR)
We measure the vulnerability as:
Impostor Attack Presentation Match Rate (IAPMR): PAs which are accepted as genuine samples – also referred to as Spoofing False Accept Rate (SFAR)
PAD sub-system: a binary classifier:
We measure 2 errors:
Attack Presentation Classification Error Rate (APCER): PAs incorrectly classified as normal presentations
Normal Presentation Classification Error Rate (NPCER): normal presentations incorrectly classified as PAs
PAD methods
- software-based: biometric data from the sensor is analysed to discriminate bona fide vs PA (eg. motion, texture)
- hardware-based: an additional sensor is used and its data analysed to discriminate bona fide vs PA (eg. temperature, pulse)
- challenge-response: the user interacts with the system (eg. prompted text in face/speaker recognition) 让用户交互起来
问
fingervein recognition?
- 指纹静脉是很有趣的,PA对这个vein是很难的啦
- 3D vein的话,让传感器建立3D结构,这个很具有挑战性,我们之前用多个相机去做过
- 这个如果数据集有的话,2D转3D很好
2D fingervein 的新的研究?
- 一个是 vein的图像非常noise,需要处理图像,预处理的算法需要搞一下,也可以借助deep model, 例如autoencoder 之类的
can adversarial attacks be considered as presentation attack?
- no
Is SWIR better for PAD? Why?
- yes
对于dynamic的可以PA么?
- 可以啦,运动的视频
- gait 是动态的,包括说话,都是可以的
- gait 可能更加困难啦
PAD是针对一个的么?
- 我们尝试让他可以泛化,可以generalize
- 看情况而定
标签:PAs,False,day5,Rate,Detection,Attack,PA,PAD,Presentation 来源: https://blog.csdn.net/MrCharles/article/details/113345866