---
tags: 生物辨識
---
# Single-Side Domain Generalization for Face Anti-Spoofing
## 目的
Poor generalization to unseen domains

> 484有點像 [這個](https://hackmd.io/Fqw4C0VmTnu9DA7xMHd9MQ)

## Contributions
- 學習真實人臉的特徵分布 (單邊)
- single-side adversarial learning 區分真假人臉
- asymmetric triplet loss 分別對真假人臉優化

## Framework

### Single-Side Adversarial Learning

> 2015 ICML Unsupervised Domain Adaptation by Backpropagation

$$
Z_{r} = G_r(X_r),\ Z_{f} = G_f(X_f)\\
\min\limits_{D}\max\limits_{G}L_{Ada}(G,D)=-E_{x,y~X_r,Y_D}\sum^N_{n=1} \unicode{x1D7D9}_{[n=y]}\log D(G(x))
$$

### Asymmetric Triplet Mining

$$
\min\limits_{G} L_{AsTrip}(G)=\sum_{x_i^\alpha,x_i^p,x_i^n}(\|f(x_i^\alpha)-f(x_i^p)\|^2-\|f(x_i^\alpha)-f(x_i^n)\|^2+\alpha)
$$

## 連結

[GITHUB](https://github.com/taylover-pei/SSDG-CVPR2020)
[Deep Learning for Face Anti-Spoofing: A Survey](https://arxiv.org/pdf/2106.14948.pdf)
[Survey GITHUB](https://github.com/ZitongYu/DeepFAS)