Wang D, Guo J, Shao Q, et al. Wild face anti-spoofing challenge 2023: Benchmark and results[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 6380-6391.
CVPRW-2023
数据集地址:https://github.com/deepinsight/insightface/tree/master/challenges/cvpr23-fas-wild
文章目录
1、Background and Motivation
背景
(1)人脸识别技术的广泛应用:
近年来,人脸识别技术已经在各个领域得到了广泛应用,如门禁控制、手机解锁、数字支付和考勤系统等。
(2)人脸识别系统的安全风险:
尽管人脸识别技术带来了便利,但这些系统也面临着显著的安全风险。例如,考勤系统可能被照片欺骗,手机解锁系统可能被攻击,数字支付系统可能被身份盗用。
(3)人脸反欺诈(FAS)技术的重要性:
人脸反欺诈(FAS)技术作为自动人脸识别系统的重要组成部分,对于保障这些系统的完整性至关重要。
(4)现有FAS数据集的局限性:
尽管FAS技术取得了显著进展,但现有方法在真实世界应用中的泛化能力仍然有限。这主要归因于公开可用的FAS数据集数量有限且缺乏多样性,导致训练时过拟合或测试时饱和。
动机
(1)解决现有数据集的局限性:
现有的FAS数据集大多包含少于2000个受试者,且样本大多在受控环境中收集,缺乏场景多样性。本文旨在通过引入一个大规模、多样化的FAS数据集来解决这些问题。
a lack of scenario richness
significant sample homogenization.
(2)引入Wild Face Anti-Spoofing (WFAS) 数据集:
WFAS数据集是在无约束环境下收集的大规模、多样化FAS数据集,包含321,751个欺诈受试者的853,729张图像和148,169个真实受试者的529,571张图像。该数据集通过互联网收集欺诈样本,涵盖了广泛的场景和各种商业传感器,包括17种呈现攻击(PAs),涵盖了2D和3D形式。
variety of scenarios and commercial sensors
(3)推动FAS技术的发展:
通过利用 WFAS 数据集和 Protocol 1(Known-Type Protocol 已知类型),本文在CVPR 2023研讨会上举办了 Wild Face Anti-Spoofing Challenge e at the CVPR2023 workshop。此外,还使用 Protocol 1 和 Protocol 2(未知类型 Unknown-Type Protocol)对代表性方法进行了细致评估。通过深入分析挑战结果和基准基线,本文提供了深入的分析,并提出了未来研究的潜在途径。
2、Related Work
-
Face Anti-Spoofing Datasets
-
Face Anti-Spoofing Methods
3、Advantages / Contributions
- 提出了大规模、多样化的 WFAS 数据集
- 基于 WFAS 提出了两种测试协议,并基于协议1 举办了 Wild Face Anti-Spoofing Challenge
- 解读分析了前三名的技术方案,给出了未来展望
4、Methods
4.1、Face Anti-Spoofing Methods
1) classification supervision
2) auxiliary pixel-wise supervision
Pseudo depth labels
Pseudo reflection maps
binary mask labels
3) generative pixel-wise supervision
Generative pixel-wise supervision is visually insightful and more interpretable.
spoof noise modeling or spoof cue generation
4.2、Wild Face Anti-Spoofing Dataset
Data Construction
obtain from the internet under a specific creative commons license,
Live Data:
All live faces were clustered using RetinaFace(人脸检测器) and ArcFace(人脸识别器),resulting in a total of 529,571 images from 148,169 live subjects.
十几万的 id,还是挺丰富的
Spoof Data:
were sourced from the internet
tablet——平板
Garage Kits——“手办”,特指需要玩家自行组装、涂装的树脂模型套件
a significant breakthrough in FAS data quantity and diversity.
training, development, and testing subsets with an approximate ratio of 4:1:5
5、Experimental Settings
5.1、Evaluation Metrics
standardized ISO/IEC 301073 metrics
Attack Presentation Classification Error Rate
Bonafide Presentation Classification Error Rate
Average Classification Error Rate
5.2、Evaluation Protocols
Known-Type Protocol——all PA types for training, development, and testing(训练、验证、测试集都包含所有的 PAs)
Unknown-Type Protocol——训练和测试的 PAs 不一样
5.3、Baselines
to eliminate the influence of arbitrary and unfaithful cues (e.g., screen bezel) on spoofing patterns,作者仅用 face 的 bbox 区域作为输入
不同颜色的框表示不同类型的方法, 依次为 classification supervision、auxiliary pixel-wise supervision、generative pixel-wise supervision
LGSC 的泛化性能不错,auxiliary pixel-wise supervision 方法不及预期
5.4、Challenge Results
Based on our dataset and Protocol 1 (Known-Type), we hosted the Wild Face Anti-Spoofing Challenge at the CVPR 2023 workshop.
The face scale was randomly set to between 1.0 and 1.2 times the side length, making it more suitable for real-world application scenarios
Netease 是网易
看看前三名的实现方案
(1)China Telecom
self-supervised stage and supervised learning stage
The teacher network’s parameters are updated with an exponential moving average (EMA) of the student parameters
老师网络和学生网络结构一样
In the second stage, self-supervised weights are used to initialize ViT for supervised learning
(2)MeiTuan
用 MoCoV2 自监督训练出 SwinV2-huge 和 SwinV2-tiny 两个网络,然后在活检数据集上微调,接着用蒸馏学习训练两个网络
(3)NetEase
先用 convnext 来 train, train 完预测一波得到 soft label
再用 maxvit 配合组合 loss 来学 soft labels
也算是蒸馏了,老师和学生的输入不同
6、Conclusion(own) / Future work
- diversity
- screen bezel
- In comparison to face recognition technology, face anti-spoofing remains an unresolved issue in face recognition systems
- the top-3 solutions all utilized the Transformer
- we have yet to see the emergence of new methods with better interpretability for the FAS task
更多论文解读,请参考 【Paper Reading】