Table 3.

Comparative analysis of techniques to mitigate face-based direct spoofing

Type/ Focus on	Ref.	Concept	Methodology Used	Dataset Used	Performance	Limitation
PS (T)	[45]	Analysis of pre-geometric and pose geometric changes through GFRPS and minimum distance classifier technique	GFRPS, MO, ED, CS, and NN classifier	PSFD	IR for local- 78.5%, and global 76.1%, GFRPS achieved 79.80% IR (Rank-1)	Hard to detect each appearance change
	[6]	A fusion of feature-based GIST and texture-based LBP methods for plastic surgery images considering edges, corners	FE- GIST, LBP, CLS- cosine distance metrics	PSFD with 1012(pre and post-surgery) (506 subjects)	VerA 91% (max)	The time complexity is not measured
	[130]	Facial marks are identified using HOG to evaluate the pre-surgery and post-surgery impact on face	Laplacian of Gaussian, HOG, SURF, Sobel, and CED	PSFD	SURF outperform others EER-42%, RR- 99.8% for FMD	Not suitable for critical plastic surgery cases
PS (T & St)	[149]	An analytical aspect is reviewed for FR system after PS with 900 individual face databases.	Polar Gabor NN transform(GNN),PCA, CLBP SURF, LFA, FDA	PSDB	GNN outperform others with 53.7% IA (Rank-1)	Sensitivity and privacy issues, not suitable to find geometric changes
	[71]	The face and ocular information-based method for face identification is proposed. A review on various surgery approaches including information of commercial software is also discussed.	FE- VJ, SIFT- LBP, Identification-Cumulative Match Characteristic	PSFD, ocular dataset	face and ocular fusion accuracy is 87.4% (Rank-1)	Low-resolution, variations in scale and expression, duplicates
	[36]	FAce Recognition against Occlusions (FARO) with expression variations divides the face into multiple regions and Partitioned Iterated Function System (PIFS) process them on the basis of codes.	FE- PCA, LDA, FARO and FACE, SFA FDA, LBP CLS- SVM, k-NN	AR-Faces	GNN algorithm’s performance among all local and global PS process.	Dataset is synthesized from benchmark DB that can affect results for real-world problem.
	[109]	A survey on state-of-the-art techniques analyzing the performance for facial plastic surgery is presented.	PCA, FDA, Geometric Features, LFA, LBP, SURF Gabor NN, PSO, PIFS, and SSIM	Public face surgery dataset	IR for GPROF method is up to 90% (rank-1)	Un-trustable high accuracy is achieved for different altered probe and gallery
	[192]	Overview of PS based FR techniques considering the relevance, applications, and surgeon’s recommendation for patient is discussed.	NA	2878 frontal images with 9 genetic disorder (36 points annotation)	Apple iPhone-X frequently updates the user’s face print information.	The nonlinear alterations are appeared in facial landmarks
	[19]	A review on contemporary research to investigate the interaction between facial PS and FR software.	FE - Circular LBP, SURF, EV-SIFT, CLS- PCA, FDA, LDA	PSFD and facial pathology.	IA in range of 15- 99%	Security, ethical, and non-linearity
	[163]	Three categories of disguise make-up (i.e., No, Light and Heavy) are investigated with considering false positive and false negative.	A three-factor (i.e., no, light and heavy) repeated-measures ANOVA is used.	24 Japanese women have participated for this reserach	RR for no makeup- high light makeup-medium heavy makeup- very low	Unable to recognize the heavy make-up.
	[34]	A non-permanent technique involving a face altering mechanism for investigating the make-up is deployed here.	Pre-processing-DoG, FE- Gabor wavelets, LBP, Projection generation-Verilook Face Toolkit, DR- PCA, LDA	YMU, VMU DB	EER range 6.50% (LBP) to 10.85% (Verilook), EER for LGBP in YMU and VMU are 15.89% (no makeup), 5.42% (full makeup)	Only female subjects are taken into consideration.
	[25]	An algorithm to classify the make-up in input image using shape, texture and color information for only considering eye and mouth features is presented.	Color space (RGB, Lab, HSV), Adaboost, SVM, GMM, LBP, DFT	YMU (151 subjects, 600 images) MIW (125 subjects, 154 images)	DR 93.5% (with 1% FPR), accuracies with SVM - 91.20 ± 0.56, with Adaboost- 89.94 ± 1.60, Overall-95.45%.	Only female subjects are considered in the database, thus not a generalized method.
	[26]	A sampling patch-based ensemble learning method to classify the image, before and after the makeup.	FE- LGGP, HGORM, DS-LBP, SRS-LDA, CLS-CRC and SRC	YMU	EERs and GARs for COTS-1, COTS-2 and COTS-3 are 12.04%, 7.69%, 9.18%, and 48.86%, 76.15%, 58.48%, respectively.	Only female subjects are considered, thus not generalized.
	[24]	A novel GAN-based unsupervised method for cycle-consistent asymmetric functions is deployed using reciprocal functions in forward (i.e., encode style transfer) and backward (i.e., destroy style) direction.	CycleGAN	Self-created dataset from YMU tutorial video	Significant results on transfer makeup styles to people having different skin-tones, original-tone with preserved identity	The performance degrades with heavy make-up.
	[148]	Disguised faces in the wild dataset is proposed with the evaluation of impersonation in three levels of difficulties (i.e., Impersonate, obfuscation, overall)	Analysis on Impersonation & obfuscation attack.	DFW, AEFRL, MiRAFace, UMDNets	IR for AEFRL of 96.80%, and Obfuscation rate for MiRA face 90.65%, and best Overall rate is 90.62% for MiRA face. (best accuracy)	The created dataset is not suitable for all types of presentation attack.

PS- Plastic Surgery, T- Textural, St- Structural, GFRPS- Geometrical Face Recognition after Plastic Surgery, Mo- Morphological operation, PSDB- Plastic Surgery Face Database, CS- Cosine similarity, NN- Neural Network, ED- Euclidean Distance, FE- Feature Extraction, CLS-Classification, EER-Equal Error rate, RR- Recognition rate, FMD- Face Mask Detection, CED- Canny Edge Detector, DoG- Difference of Gaussian, CLBP- Circular Local Binary Pattern,IA- Identification Accuracy, GNN- Neural Network-based Gabor Transform, SFA- Split Face Architecture, YMU- YouTube MakeUp,VMU- Synthetic Virtual Makeup,LGGP- Local Gradient Gabor Pattern, HGORMHistogram of Gabor Ordinal Ratio Measures, DS- Densely Sampled, CRC-Collaborativebased Representation Classifiers, SRC- Sparse-based Representation Classifiers, PSO- Particle Swarm Optimization, PIFS- Partitioned Iterated Function System, SSIM-Structural Similarity Image Maps