---
tags: 生物辨識
---
# Generalized Presentation Attack Detection: a face anti-spoofing evaluation proposal
制定一個統一的標準去評估人臉防偽模型的泛化性
> 現在大家的方法幾乎都會綁定到特定的資料集上面
- 設計實驗模擬真實場景。
-
## Contributions
- Largest dataset
- anti-spoofing: attacks, lighting, capture devices and resolution.
- We release an open-source evaluation framework.
- We demonstrate the limitation of current dataset evaluation procedures.
- We introduce two novel protocols for the GPAD problem.
1. Cross-FaceResolution
2. Cross-Conditions

## Generalized Presentation Attack Detection
Regardless almost every proposal comes with its own reduced dataset, there is no agreement upon a PAD benchmark, and generalization properties are not properly evaluated
> 現在最直接的做法都推薦直接採用不同dataset去驗證泛化性


- Stage1: Feature Extraction
- Stage2: Evaluation

> PAI: Presentation Attack Instruments

- The Aggregated Dataset


- Categorization
> 對各種 attck/capture devices/lighting/face resolution 去做額外的分類。


- Protocols（考慮到各種情境）
1. Grandtest (without any filter)
2. Cross-Dataset (直接測試在不同的資料集)
3. One-PAI
4. Unseen Attacks (Cross-PAI)
5. Unseen Capture Devices
6. **Cross-FaceResolution**：高解析度訓練，低解析度測試
7. **Cross-Conditions**：在有利的情況以及不利的情況下去做比較


## 連結
- [Paper](https://arxiv.org/pdf/1904.06213.pdf)
- [Github](https://github.com/acostapazo/gradgpad)