---
tags: 生物辨識
---
# CelebA-Spoof Challenge 2020 on Face Anti-Spoofing: Methods and Results

## Dataset
- Live Data
- The live data are directly inherited from CelebA dataset
- Manually examine the images in CelebA and remove those “spoof” images, including posters, advertisements and cartoon portrait
- Spoof Data
- We hired 8 collectors to collect spoof data and another 2 annotators to refine labeling for all data.
- Attribute
- 40 types of Face Attribute defined in CelebA plus 3 attributes of face anti-spoofing, including Spoof Type, Illumination Condition, and Environment.

## AE-Net(Baseline)
- Semantic information
- Semantic for Live
- $S^f$ : Face Attribute
- Semantic for spoof
- $S^s$ : Spoof Type
- $S^i$ : Illumination Type
- $C$ : live/spoof class
- geometric information
- $g^d$ : depth map
- $g^r$ : reflection map

## Third place
- Method
- Train with Patches
- Ensemble
- 5 model with different input size
- Light-weighted Network1 : input size 32x32
- Light-weighted Network2 : input size 48x48
- CDCN : input size 64x64
- CDC-DAN (self-Attention CDCN) : input size 112x112
- Se-resnext26: input size 128x128
- fusion strategy
- fusion strategy is a simple method that adjusts the weights of CNN models based on the best performance of the validation dataset
- Training Strategy
- Data augmentation(cutout, vh-mixup, mixed-concat, random square and random interval)
- if the image size is smaller than the preset input size, they enlarge the image by mirroring instead of scaling, which greatly improves the performance.
- Testing Strategy
- They only adjust the image size by mirroring to meet different network input requirements.

## Second Place
- Method
- Patches
- Ensemble
- 5 model with different input size
- $CDCN_{pp}$ : trained the CDCNpp on two scales of patches, 64x64 and 96x96
- LGSC : reisze to 224x224
- SeResNet50 : reisze to 224x224
- EfficientNet-b7: reisze to 224x224
- SeResNeXt50
- random patches to 64*64
- Multi-task learning, add 2 fc layer in the tail of SeResNeXt50, predicting the spoof types and the illumination types respectively.
- fusion strategy
- weighting-after-sorting
- 將六個輸出的score依大至小排序, 然後取出前四大的score乘上相對應的weight當作最後輸出的分數
- Use Particle Swarm Optimization (PSO) algorithm to find the k weights assigned to the top k scores at different ranks with the best performance on the validation set.
- Testing Strategy
- $CDCN_{pp}$
1. Image was first split into 3*3 parts.
2. Crop the upper left corner with size of 64x64 and the lower right corner with size of 96x96 of each part.
- SeResNeXt50
- Crop the center part with size of 64*64 of the image and flip the patch horizontally
- others : directly resize image

## First place
- Method
- Ensemble
- 5 model
- FOCUS(Finding spOof CUe for face anti-Spoofing)
- AENet
- ResNet
- focal loss
- data augmentation
- Attack type classification
- 發現不同spoof的照片有相同的攻擊線索
- display borders
- similar backgrounds
- similar paper printing edges
- 將人臉部分去除掉之後做影像分類
- Noise Print
- live images are collected from internet or social media
- Spoof images are directly captured from device cameras, e.g. Phone camera, Pad camera or PC camera.
- acquired images from different cameras have artifacts which are peculiar to the camera itself
- Classify 4 group images (live, Phone, Pad, PC)
- fusion strategy(為了得到最好的TAR@FAR)
1. 標準化每個model的confidence scores至0-1
2. 讓在Validation裡表現最好的model當作main model, 其他的model當作輔助的model
3. 如果每一個model都有一樣的預測結果的話, 將最後輸出的confidence score變成0或1
4. 如果輔助的model的confidence score很大的話, 將最後輸出的結果換成輔助的model結果,然後把confidence score變成0或1
5. 將所有model預測出來的score都不接近0和1的image視為hard case, 並將他最後輸出的confidence score變成0.1