👏 Survey of Deep Face Anti-spoofing 🔥

This is the official repository of "Deep Learning for Face Anti-Spoofing: A Survey", a comprehensive survey of recent progress in deep learning methods for face anti-spoofing (FAS) as well as the datasets and protocols.

Introduction

We present a comprehensive review of recent deep learning methods for face anti-spoofing (mostly from 2018 to 2021). It covers hybrid (handcrafted+deep), pure deep learning, and generalized learning based methods for monocular RGB face anti-spoofing. It also includes multi-modal learning based methods as well as specialized sensor based FAS. It also presents detailed comparision among publicly available datasets, together with several classical evaluation protocols.

🔔 We will update this page frequently~ 🎉🎉🎉

Dataset	Year	#Live/Spoof	#Sub.	Setup	Attack Types
NUAA	2010	5105/7509(I)	15	N/R	Print(flat, wrapped)
YALE Recaptured	2011	640/1920(I)	10	50cm-distance from 3 LCD minitors	Print(flat)
CASIA-MFSD	2012	150/450(V)	50	7 scenarios and 3 image quality	Print(flat, wrapped, cut), Replay(tablet)
REPLAY-ATTACK	2012	200/1000(V)	50	Lighting and holding	Print(flat), Replay(tablet, phone)
Kose and Dugelay	2013	200/198(I)	20	N/R	Mask(hard resin)
MSU-MFSD	2014	70/210(V)	35	Indoor scenario; 2 types of cameras	Print(flat), Replay(tablet, phone)
UVAD	2015	808/16268(V)	404	Different lighting, background and places in two sections	Replay(monitor)
REPLAY-Mobile	2016	390/640(V)	40	5 lighting conditions	Print(flat), Replay(monitor)
HKBU-MARs V2	2016	504/504(V)	12	7 cameras from stationary and mobile devices and 6 lighting settings	Mask(hard resin) from Thatsmyface and REAL-f
MSU USSA	2016	1140/9120(I)	1140	Uncontrolled; 2 types of cameras	Print(flat), Replay(laptop, tablet, phone)
SMAD	2017	65/65(V)	-	Color images from online resources	Mask(silicone)
OULU-NPU	2017	720/2880(V)	55	Lighting & background in 3 sections	Print(flat), Replay(phone)
Rose-Youtu	2018	500/2850(V)	20	5 front-facing phone camera; 5 different illumination conditions	Print(flat), Replay(monitor, laptop),Mask(paper, crop-paper)
SiW	2018	1320/3300(V)	165	4 sessions with variations of distance, pose, illumination and expression	Print(flat, wrapped), Replay(phone, tablet, monitor)
WFFD	2019	2300/2300(I) 140/145(V)	745	Collected online; super-realistic; removed low-quality faces	Waxworks(wax)
SiW-M	2019	660/968(V)	493	Indoor environment with pose, lighting and expression variations	Print(flat), Replay, Mask(hard resin, plastic, silicone, paper, Mannequin), Makeup(cosmetics, impersonation, Obfuscation), Partial(glasses, cut paper)
Swax	2020	Total 1812(I) 110(V)	55	Collected online; captured under uncontrolled scenarios	Waxworks(wax)
CelebA-Spoof	2020	156384/469153(I)	10177	4 illumination conditions; indoor & outdoor; rich annotations	Print(flat, wrapped), Replay(monitor tablet, phone), Mask(paper)
RECOD-Mtablet	2020	450/1800(V)	45	Outdoor environment and low-light & dynamic sessions	Print(flat), Replay(monitor)
CASIA-SURF 3DMask	2020	288/864(V)	48	High-quality identity-preserved; 3 decorations and 6 environments	Mask(mannequin with 3D print)
HiFiMask	2021	13650/40950(V)	75	three mask decorations; 7 recording devices; 6 lighting conditions; 6 scenes	Mask(transparent, plaster, resin)

Datasets with multiple modalities or specialized sensors

Dataset	Year	#Live/Spoof	#Sub.	M&H	Setup	Attack Types
3DMAD	2013	170/85(V)	17	VIS, Depth	3 sessions (2 weeks interval)	Mask(paper, hard resin)
GUC-LiFFAD	2015	1798/3028(V)	80	Light field	Distance of 1.5 constrained conditions	Print(Inkjet paper, Laserjet paper), Replay(tablet)
3DFS-DB	2016	260/260(V)	26	VIS, Depth	Head movement with rich angles	Mask(plastic)
BRSU Skin/Face/Spoof	2016	102/404(I)	137	VIS, SWIR	multispectral SWIR with 4 wavebands 935nm, 1060nm, 1300nm and 1550nm	Mask(silicon, plastic, resin, latex)
Msspoof	2016	1470/3024(I)	21	VIS, NIR	7 environmental conditions	Black&white Print(flat)
MLFP	2017	150/1200(V)	10	VIS, NIR, Thermal	Indoor and outdoor with fixed and random backgrounds	Mask(latex, paper)
ERPA	2017	Total 86(V)	5	VIS, Depth, NIR, Thermal	Subject positioned close (0.3∼0.5m) to the 2 types of cameras	Print(flat), Replay(monitor), Mask(resin, silicone)
LF-SAD	2018	328/596(I)	50	Light field	Indoor fix background, captured by Lytro ILLUM camera	Print(flat, wrapped), Replay(monitor)
CSMAD	2018	104/159(V+I)	14	VIS, Depth, NIR, Thermal	4 lighting conditions	Mask(custom silicone)
3DMA	2019	536/384(V)	67	VIS, NIR	48 masks with different ID; 2 illumination & 4 capturing distances	Mask(plastics)
CASIA-SURF	2019	3000/18000(V)	1000	VIS, Depth, NIR	Background removed; Randomly cut eyes, nose or mouth areas	Print(flat, wrapped, cut)
WMCA	2019	347/1332(V)	72	VIS, Depth, NIR, Thermal	6 sessions with different backgrounds and illumination; pulse data for bonafide recordings	Print(flat), Replay(tablet), Partial(glasses), Mask(plastic, silicone, and paper, Mannequin)
CeFA	2020	6300/27900(V)	1607	VIS, Depth, NIR	3 ethnicities; outdoor & indoor; decoration with wig and glasses	Print(flat, wrapped), Replay, Mask(3D print, silica gel)
HQ-WMCA	2020	555/2349(V)	51	VIS, Depth, NIR, SWIR, Thermal	Indoor; 14 ‘modalities’, including 4 NIR and 7 SWIR wavelengths; masks and mannequins were heated up to reach body temperature	Laser or inkjet Print(flat), Replay(tablet, phone), Mask(plastic, silicon, paper, mannequin), Makeup, Partial(glasses, wigs, tatoo)

2️⃣ Deep FAS methods with commercial RGB camera

temp

Hybrid (handcrafted + deep)

Method	Year	Backbone	Loss	Input	Static/Dynamic
DPCNN	2016	VGG-Face	Trained with SVM	RGB	S
Multi-cues+NN	2016	MLP	Binary CE loss	RGB+OFM	D
CNN LBP-TOP	2017	5-layer CNN	Binary CE loss, SVM	RGB	D
DF-MSLBP	2018	Deep forest	Binary CE loss	HSV+YCbCr	S
SPMT+SSD	2018	VGG16	Binary CE loss, SVM, bbox regression	RGB, Landmarks	S
CHIF	2019	VGG-Face	Trained with SVM	RGB	S
DeepLBP	2019	VGG-Face	Binary CE loss, SVM	RGB, HSV, YCbCr	S
CNN+LBP+WLD	2019	CaffeNet	Binary CE loss	RGB	S
Intrinsic	2019	1D-CNN	Trained with SVM	Reflection	D
FARCNN	2019	Multi-scale attentional CNN	Regression loss, Crystal loss, Center loss	RGB	S
CNN-LSP	TIFS 2019	1D-CNN	Trained with SVM	RGB	D
DT-Mask	2019	VGG16	Binary CE loss, Channel&Spatial discriminability	RGB+OF	D
VGG+LBP	2019	VGG16	Binary CE loss	RGB	S
CNN+OVLBP	2019	VGG16	Binary CE loss, NN classifier	RGB	S
HOG-Pert.	2019	Multi-scale CNN	Binary CE loss	RGB+HOG	S
LBP-Pert.	2020	Multi-scale CNN	Binary CE loss	RGB+LBP	S
TransRPPG	SPL 2021	Vision Transformer	Binary CE loss	rPPG map	D

End-to-end binary cross-entropy supervision

Method	Year	Backbone	Loss	Input	Static/Dynamic
CNN1	2014	8-layer CNN	Trained with SVM	RGB	S
LSTM-CNN	2015	CNN+LSTM	Binary CE loss	RGB	D
SpoofNet	2015	2-layer CNN	Binary CE loss	RGB	S
HybridCNN	2017	VGG-Face	Trained with SVM	RGB	S
CNN2	2017	VGG11	Binary CE loss	RGB	S
Ultra-Deep	2017	ResNet50+LSTM	Binary CE loss	RGB	D
FASNet	2017	VGG16	Binary CE loss	RGB	S
CNN3	2018	Inception, ResNet	Binary CE loss	RGB	S
MILHP	2018	ResNet+STN	Multiple Instances CE loss	RGB	D
LSCNN	2018	9 PatchNets	Binary CE loss	RGB	S
LiveNet	2018	VGG11	Binary CE loss	RGB	S
MS-FANS	2018	AlexNet+LSTM	Binary CE loss	RGB	S
DeepColorFAS	2018	5-layer CNN	Binary CE loss	RGB, HSV, YCbCr	S
Siamese	2019	AlexNet	Contrastive loss	RGB	S
FSBuster	2019	ResNet50	Trained with SVM	RGB	S
FuseDNG	2019	7-layer CNN	Binary CE loss, Reconstruction loss	RGB	S
STASN	CVPR 2019	ResNet50+LSTM	Binary CE loss	RGB	D
TSCNN	TIFS 2019	ResNet18	Binary CE loss	RGB, MSR	S
FAS-UCM	2019	MobileNetV2, VGG19	Binary CE loss, Style loss	RGB	S
SLRNN	2019	ResNet50+LSTM	Binary CE loss	RGB	D
GFA-CNN	2019	VGG16	Binary CE loss	RGB	S
3DSynthesis	2019	ResNet15	Binary CE loss	RGB	S
CompactNet	NC 2020	VGG19	Points-to-Center triplet loss	RGB	S
SSR-FCN	TIFS 2020	FCN with 6 layers	Binary CE loss	RGB	S
FasTCo	2020	ResNet50 or MobileNetV2	Multi-class CE loss, Temporal Consistency loss, Class Consistency loss	RGB	D
DRL-FAS	TIFS 2020	ResNet18+GRU	Binary CE loss	RGB	S
SfSNet	2020	6-layer CNN	Binary CE loss	Albedo, Depth, Reflection	S
LivenesSlight	2020	6-layer CNN	Binary CE loss	RGB	S
MotionEnhancement	2020	VGGface+LSTM	Binary CE loss	RGB	D
CFSA-FAS	2020	ResNet18	Binary CE loss	RGB	S
MC-FBC	2020	VGG16, ResNet50	Binary CE loss	RGB	S
SimpleNet	2020	Multi-stream 5-layer CNN	Binary CE loss	RGB, OF, RP	D
PatchCNN	2020	SqueezeNet v1.1	Binary CE loss, Triplet loss	RGB	S
FreqSpatialTempNet	2020	ResNet18	Binary CE loss	RGB, HSV, Spectral	D
ViTranZFAS	2020	Vision Transformer	Binary CE loss	RGB	S
CIFL	TIFS 2021	ResNet18	Binary focal loss, camear type loss	RGB	S

Pixel-wise auxiliary supervision

Method	Year	Supervision	Backbone	Input	Static/Dynamic
Depth&Patch	IJCB 2017	Depth	PatchNet, DepthNet	YCbCr, HSV	S
Auxiliary	CVPR 2018	Depth, rPPG spectrum	DepthNet	RGB, HSV	D
BASN	ICCVW 2019	Depth, Reflection	DepthNet, Enrichment	RGB, HSV	S
DTN	CVPR 2019	BinaryMask	Tree Network	RGB, HSV	S
PixBiS	ICB 2019	BinaryMask	DenseNet161	RGB	S
A-PixBiS	2020	BinaryMask	DenseNet161	RGB	S
Auto-FAS	ICASSP 2020	BinaryMask	NAS	RGB	S
MRCNN	2020	BinaryMask	Shallow CNN	RGB	S
FCN-LSA	2020	BinaryMask	DepthNet	RGB	S
CDCN	CVPR 2020	Depth	DepthNet	RGB	S
FAS-SGTD	CVPR 2020	Depth	DepthNet, STPM	RGB	D
TS-FEN	2020	Depth	ResNet34, FCN	RGB, YCbCr, HSV	S
SAPLC	2020	TernaryMap	DepthNet	RGB, HSV	S
BCN	ECCV 2020	BinaryMask, Depth, Reflection	DepthNet	RGB	S
Disentangled	ECCV 2020	Depth, TextureMap	DepthNet	RGB	S
AENet	ECCV 2020	Depth, Reflection	ResNet18	RGB	S
3DPC-Net	2020	3D Point Cloud	ResNet18	RGB	S
PS	TBIOM 2020	BinaryMask or Depth	ResNet50 or CDCN	RGB	S
NAS-FAS	PAMI 2020	BinaryMask or Depth	NAS	RGB	D
DAM	2021	Depth	VGG16, TSM	RGB	D
Bi-FPNFAS	2021	Fourier spectra	EfficientNetB0, FPN	RGB	S
DC-CDN	IJCAI 2021	Depth	CDCN	RGB	S

Generative model with pixel-wise supervision

Method	Year	Supervision	Backbone	Input	Static/Dynamic
De-Spoof	ECCV 2018	Depth, BinaryMask, FourierMap	DSNet, DepthNet	RGB, HSV	S
Reconstruction	2019	RGB Input (live), ZeroMap (spoof)	U-Net	RGB	S
LGSC	2020	ZeroMap (live)	U-Net, ResNet18	RGB	S
TAE	ICASSP 2020	Binary CE loss, Reconstruction loss	Info-VAE, DenseNet161	RGB	S
STDN	ECCV 2020	BinaryMask, RGB Input (live)	U-Net, PatchGAN	RGB	S
GOGen	CVPR 2020	RGB input	DepthNet	RGB+one-hot vector	S
PhySTD	2021	Depth, RGB Input (live)	U-Net, PatchGAN	Frequency Trace	S
MT-FAS	PAMI 2021	ZeroMap (live), LearnableMap (Spoof)	DepthNet	RGB	S

Domain adaptation

Method	Year	Backbone	Loss	Static/Dynamic
OR-DA	TIFS 2018	AlexNet	Binary CE loss, MMD loss	S
DTCNN	2019	AlexNet	Binary CE loss, MMD loss	S
Adversarial	ICB 2019	ResNet18	Triplet loss, Adversarial loss	S
ML-MMD	ICMEW 2019	Multi-scale FCN	CE loss, MMD loss	S
OCA-FAS	NC 2020	DepthNet	Binary CE loss, Pixel-wise binary loss	S
DR-UDA	TIFS 2020	ResNet18	Center&Triplet loss, Adversarial loss, Disentangled loss	S
DGP	ICASSP 2020	DenseNet161	Feature divergence measure, BinaryMask loss	S
Distillation	J-STSP 2020	AlexNet	Binary CE loss, MMD loss , Paired Similarity	S
SCNN++PL+TC	TIP 2021	ResNet18	CE Loss in labeled and unlabeled sets	D
USDAN	PR 2021	ResNet18	Adaptive binary CE loss, Entropy loss, Adversarial loss	S
SASA	2021	ResNet18	CE Loss, Adversarial loss, Less-forgetting constraints, Contrastive semantic alignment	S

Domain generalization

Method	Year	Backbone	Loss	Static/Dynamic
MADDG	CVPR 2019	DepthNet	Binary CE & Depth loss, Multi-adversarial loss, Dual-force Triplet loss	S
PAD-GAN	CVPR 2020	ResNet18	Binary CE & Depth loss, Multi-adversarial loss, Dual-force Triplet loss	S
DASN	2020	ResNet18	Binary CE & Spoof-irrelevant factor loss	S
SSDG	CVPR 2020	ResNet18	Binary CE loss, Single-Side adversarial loss, Asymmetric Triplet loss	S
RF-Meta	AAAI 2020	DepthNet	Binary CE loss, Depth loss	S
CCDD	CVPRW 2020	ResNet50+LSTM	Binary CE loss, Class-conditional loss	D
SDA	AAAI 2021	DepthNet	Binary CE & Depth loss, Reconstruction loss, Orthogonality regularization	S
D2AM	AAAI 2021	DepthNet	Binary CE loss, Depth loss, MMD loss	S
DRDG	IJCAI 2021	DepthNet	Binary CE loss, Depth loss, Domain loss	S

Zero/Few-shot learning

Method	Year	Backbone	Loss	Input
DTN	CVPR 2019	Deep Tree Network	Binary CE loss, Pixel-wise binary loss, Unsupervised Tree loss	RGB, HSV
AIM-FAS	AAAI 2020	DepthNet	Depth loss, Contrastive Depth loss	RGB
CM-PAD	IJCB 2021	DepthNet, ResNet	Binary CE loss, Depth loss, Gradient alignment	RGB

Anomaly detection

Method	Year	Backbone	Loss	Input
AE+LBP	2018	AutoEncoder	Reconstruction loss	RGB
Anomaly	2019	ResNet50	Triplet focal loss, Metric-Softmax loss	RGB
Anomaly2	2019	GoogLeNet or ResNet50	Mahalanobis distance	RGB
Hypersphere	2020	ResNet18	Hypersphere loss	RGB, HSV
Ensemble-Anomaly	2020	GoogLeNet or ResNet50	Gaussian Mixture Model (not end-to-end)	RGB, patches
MCCNN	2020	LightCNN	Binary CE loss, Contrastive loss	Grayscale, IR, Depth, Thermal
End2End-Anomaly	2020	VGG-Face	Binary CE loss, Pairwise confusion	RGB
ClientAnomaly	PR 2020	ResNet50 or GoogLeNet or VGG16	One-class SVM or Mahalanobis distance or Gaussian Mixture Model	RGB

3️⃣ Deep FAS methods with advanced sensor

Learning upon specialized sensor

Method	Year	Backbone	Loss	Input	Static/Dynamic
Thermal-FaceCNN	2019	AlexNet	Regression loss	Thermal infrared face image	S
SLNet	2019	17-layer CNN	Binary CE loss	Stereo (left&right) face images	S
Aurora-Guard	2019	U-Net	Binary CE loss, Depth regression, Light Regression	Casted face with dynamic changing light specified by random light CAPTCHA	D
LFC	2019	AlexNet	Binary CE loss	Ray difference/microlens images from light field camera	S
PAAS	2020	MobileNetV2	Contrastive loss, SVM	Four-directional polarized face image	S
Face-Revelio	2020	Siamese-AlexNet	L1 distance	Four flash lights displayed on four quarters of a screen	D
SpecDiff	2020	ResNet4	Binary CE loss	Concatenated face images w/ and w/o flash	S
MC-PixBiS	2020	DenseNet161	Binary mask loss	SWIR images differences	S
Thermalization	2020	YOLO V3+GoogLeNet	Binary CE loss	Thermal infrared face image	S
DP Bin-Cls-Net	2021	Shallow U-Net + Xception	Transformation consistency, Relative disparity loss, Binary CE loss	DP image pair	S

Multi-modal learning

Method	Year	Backbone	Loss	Input	Fusion
FaceBagNet	2019	Multi-stream CNN	Binary CE loss	RGB, Depth, NIR face patches	Feature-level
FeatherNets	2019	Ensemble-FeatherNet	Binary CE loss	Depth, NIR	Decision-level
Attention	2019	ResNet18	Binary CE loss, Center loss	RGB, Depth, NIR	Feature-level
mmfCNN	ACMMM 2019	ResNet34	Binary CE loss, Binary Center Loss	RGB, NIR, Depth, HSV, YCbCr	Feature-level
MM-FAS	2019	ResNet18/50	Binary CE loss	RGB, NIR, Depth	Feature-level
AEs+MLP	2019	Autoencoder, MLP	Binary CE loss, Reconstruction loss	Grayscale-Depth-Infrared composition	Input-level
SD-Net	2019	ResNet18	Binary CE loss	RGB, NIR, Depth	Feature-level
Dual-modal	2019	MoblienetV3	Binary CE loss	RGB, IR	Feature-level
Parallel-CNN	2020	Attentional CNN	Binary CE loss	Depth, NIR	Feature-level
Multi-Channel Detector	2020	RetinaNet (FPN+ResNet18)	Landmark regression, Focal loss	Grayscale-Depth-Infrared composition	Input-level
PSMM-Net	2020	ResNet18	Binary CE loss for each stream	RGB, Depth, NIR	Feature-level
PipeNet	2020	SENet154	Binary CE loss	RGB, Depth, NIR face patches	Feature-level
MM-CDCN	2020	CDCN	Pixel-wise binary loss, Contrastive depth loss	RGB, Depth, NIR	Feature&Decision-level
HGCNN	2020	Hypergraph-CNN, MLP	Binary CE loss	RGB, Depth	Feature-level
MCT-GAN	2020	CycleGAN, ResNet50	GAN loss, Binary CE loss	RGB, NIR	Input-level
D-M-Net	2021	ResNeXt	Binary CE loss	Multi-preprocessed Depth, RGB-NIR composition	Input&Feature-level
CMFL	CVPR 2021	DenseNet161	Binary CE loss, Cross modal focal loss	RGB, Depth	Feature-level
MA-Net	TIFS 2021	CycleGAN, ResNet18	Binary CE loss, GAN loss	RGB, NIR	Feature-level

Citation

If you find our work useful in your research, please consider citing:

@article{yu2021deep,
  title={Deep Learning for Face Anti-Spoofing: A Survey},
  author={Yu, Zitong and Qin, Yunxiao and Li, Xiaobai and Zhao, Chenxu and Lei, Zhen and Zhao, Guoying},
  journal={arXiv},
  year={2021}
}

About

🔥Deep Learning for Face Anti-Spoofing

chaoso / DeepFAS