Table 4.

Comparative analysis of techniques to mitigate face-based indirect spoofing

Type/ Focus on	Ref.	Concept	Methodology Used	Dataset Used	Performance	Limitation
PA	[29]	A new partial least squares-discriminant analysis method to reduce the modularity impact in visible and thermal images	LBP, HOG, Gabor filter and PLSDA	(X1 Collection) UND dataset	Thermal-visible Facial IR 49.9%.	Modality gap is more, less efficient to night vision
	[51]	A novel overlapping digital grid method to detect the presentation attack by utilizing Moiré pattern is proposed.	FE- Moiré Pattern digital grid, CLS- SVM (RBF kernel)	MIT-CBCL, Yale, Caltech	This method provides good results for low pixel ratio.	Not suitable with illumination variance problem
	[20]	A non-intrusive software-based method to evaluate the luminance color information from the face image.	Disparity, CoALBP, LBP, LPQ, BSIF, SID, SVM with histogram	CASIA-FASD, Replay Attack, MSU-MFSD	EER For given DB (in %)-Video- 3.2, 0, 3.5, Image- 2.1, 0.4, 4.9, respectively.	The color-texture analysis is not generalized.
	[138]	The Light field disparity-based face PAD techniques are reviewed with introducing a new HOG descriptor.	HOG, SVM	NUAA, Print attack, Replay Attack, CASIA	ACER- Laptop 0.05%, Mobile 0.02% paper -0.5%	Unavailability of appropriate Light-field DB, focus- based poor results
	[27]	A CNN-based method with color-texture features (RI-LBP) is introduced here.	FE- ResNet-18 and rotation invariant -LBP), CLS- SVM, DR- PCA	NUAA Replay CASIA-FASD, MSU-MFSD	ERR for NUAA, Replay, CASIA, MSU are 0.5, 2.3, 4.4, 3.1 respectively.	The results on cross-DB must be generalized.
	[124]	A new database SCFASD is introduced with disparity layer-based supervised CNN classifier to efficiently detect face liveness.	Dynamic disparity maps and CNN	Stereo face anti-spoofing DB for (printed photo, mobile, tablet)	Overall APCER = 0.86 ± 0.80	More generalization is expected for real-time applications.
	[56]	The adaptive fusion of each classifier scores is performed on multimodal biometrics (face, finger, and iris) considering the concurrent (boosted) and discordant (suppressed) Modalities.	Hough Transform, Gabor, minutiae features, Gaussian kernel function, Min-max, threshold	Chimeric multimodal databases	Average accuracy 99.5%, EER 0.5%	It doesn’t support a dynamic environment
	[123]	DL-based facial PAD method using perturbation with pre-processing is proposed.	CNN, LBP	CASIA, Idiap Replay-Attack, OULU-NPU DB.	ACER (in %)- 3.89 (Oulu NPU), 0.23 (CASIA and Idiap), 0.97 (MSU-USSA)	Other handcrafted method also has to be evaluated.
VA	[84]	A LBP and CNN-based methods are comparatively evaluated for face spoofing.	LBP and CNN	Replay Attack and CASIA-FA	ACER- Replay Attack 1.3%, CASIA-FA 2.5%	Slow speed recognition, sensitive to noise
	[151]	Three discriminative (SPMT, SSD, TFBD) representations for face PAD is performed.	FE- SPMT, SSD, TFBD, CLS- SVM	Replay Attack and CASIA-FA	ACER- Replay Attack 0%, CASIA-FD- 2.58%	Sensitive to stereo type, binocular camera is needed
	[178]	A deep architecture-based FLD to prevent video spoofing attack is proposed here.	FE-CNN and generalized multiple kernel learning (GMKL),CLS- SVM	Replay Attack and CASIA-FA	Replay Attack 0% CASIA-FA 2.58%	Edge enhancement, texture differences are not considered.
	[69]	Several face PAD techniques are reviewed to generalize the mobile-based authentication for various cross-databases.	Mobile spoofing datasets	MSU-USSA, Replay Mobile, Oulu NPU	ACER (in %)- 8.33 (Oulu NPU), 0.26 (Replay mobile), 0.97 (MSU-USSA)	The biased DB has to be more practical and generalized.
MA	[121]	A comprehensive review for the state-of-the-art anti spoofing techniques are investigated here.	Challenge response, blink detection	Replay Attack, 3D mask Attack, Print Attack DB	APCER for Print Wrap and display are 5.27%, 1.21%, 0.71%, respectively	H/W: costly, overhead, S/W: device dependent, Not generalized
	[52]	A multi-channel CNN method with a new WMCA dataset consisting of 2D and 3D attack for Impersonation and Obfuscation condition is introduced here.	CNN with a novel Wide Multi-Channel PAD	WMCA dataset with thermal, near-infrared, color, and depth information.	ACER of 0.3%	Unseen attack protocol evaluation for problems is not efficient.
	[49]	An image-quality assessment-based fast non-intrusive method with motion cues is proposed for face spoofing detection.	LDA, SVM	Replay attack, Replay mobile, 3D MAD	HTER-0.024% (Replay attack)	This process is not suitable for large amount of database.

PA-Photo attack, VA- video attack, MA- Mask attack, PAD- Presentation attack detection, FLD- face liveness detection, SPMT- spatial pyramid coding micro-texture, TFBD-template face matched binocular depth, SSD- single shot multiBox detector,IR- Identification Rate, APCER-Attack presentation classification error rate, ACER- Average classification error rate, HTER- Half total error rate, WMCPA-Wide Multi-Channel Presentation Attack dataset