Javascript is required
1.
B. Choudhury, M. Haldar, B. Issac, V. Raman, and P. Then, “A survey on biometrics and cancelable biometrics systems,” Int. J. Image Graph., vol. 18, no. 1, pp. 1–39, 2018. [Google Scholar] [Crossref]
2.
D. Kalenichenko, J. Philbin, and F. Schroff, “Facenet: A unified embedding for face recognition and clustering,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, June 07-12, 2015. pp. 815–823. [Google Scholar] [Crossref]
3.
C. Busch and R. Ramachandra, “Presentation attack detection methods for face recognition systems: A comprehensive survey,” ACM Comput. Surv., vol. 50, no. 1, pp. 1–37, 2017. [Google Scholar] [Crossref]
4.
R. H. Deng, Y. Li, K. Xu, Y. J. Li, and Q. Yan, “Understanding OSN-based facial disclosure against face authentication systems,” Proceedings of the 9th ACM Symposium on Information, Computer and Communications Security, Kyoto, Japan, June 4-6, 2014. pp. 413–424. [Google Scholar] [Crossref]
5.
X. Y. Tan, Y. Li, J. Liu, and L. Jiang, “Face liveness detection from a single image with sparse low rank bilinear discriminative model,” Computer Vision – ECCV 2010. ECCV 2010. Lecture Notes in Computer Science. Springer, Berlin, Heidelberg, pp. 504–517, 2010. [Google Scholar] [Crossref]
6.
A. Orjiude, “Eight banks lose N1.9bn to fraud in one year,” Punch Newspaper, 2021. https://punchng.com/eight-banks-lose-n1-9bn-to-fraud-in-one-year/ [Google Scholar]
7.
Z. F. Li, Y. Qiao, K. P. Zhang, and Z. P. Zhang, “Joint face detection and alignment using multitask cascaded convolutional networks,” IEEE Signal Process. Lett., vol. 23, no. 10, pp. 1499–1503, 2016. [Google Scholar] [Crossref]
8.
Z. W. Zhang, J. J. Yan, S. F. Liu, Z. Li, Z. Lei, D. Yi, and S. Z. Li, “A face anti-spoofing database with diverse attacks,” in 5th IAPR International Conference on Biometrics (ICB), New Delhi, India, March 29, 2012-April 01, 2012, pp. 26–31. [Google Scholar] [Crossref]
9.
T. Mäenpää, T. Ojala, and M. Pietikäinen, “Multiresolution grayscale and rotation invariant texture classification with local binary patterns,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 7, pp. 971–987, 2002. [Google Scholar] [Crossref]
10.
J. Komulainen, A. Hadid, and M. Pietikäinen, “Face spoofing detection from single images using texture and local shape analysis,” IET Biometrics, vol. 1, no. 1, pp. 3–10, 2012. [Google Scholar] [Crossref]
11.
N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” Proceedings of IEEE Computing Society Conference on Computer Vision and Pattern Recognition, San Diego, CA, USA. pp. 886–893, 2005. [Google Scholar] [Crossref]
12.
B. S. Manjunath and W. Y. Ma, “Texture features for browsing and retrieval of image data,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 18, no. 8, pp. 837–842, 1996. [Google Scholar] [Crossref]
13.
D. Gragnaniello, G. Poggi, C. Sansone, and L. Verdoliva, “An investigation of local descriptors for biometric spoofing detection,” IEEE Trans. Inf. Forensics Secur., vol. 10, no. 4, pp. 849–863, 2015. [Google Scholar] [Crossref]
14.
H. X. Li, Z. Lin, X. H. Shen, J. Brandt, and G. Hua, “A convolutional neural network cascade for face detection,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, June 07-12, 2015. pp. 5325–5334. [Google Scholar] [Crossref]
15.
L. Li, X. Y. Feng, X. Y. Jiang, Z. Q. Xia, and A. Hadid, “Face anti-spoofing via deep local binary patterns,” Proceedings of IEEE International Conference on Image Processing, Beijing, China, September 17-20, 2017. pp. 101–105. [Google Scholar] [Crossref]
16.
K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint, 2014. [Google Scholar] [Crossref]
17.
O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and F. F. Li, “ImageNet large scale visual recognition challenge,” Int. J. Comput. Vis., vol. 115, no. 3, pp. 211–252, 2015. [Google Scholar] [Crossref]
18.
Z. Zhang, C. Jiang, X. Zhong, C. Song, and Y. Zhang, “Two-stream convolutional networks for multi-frame face anti-spoofing,” arXiv preprint, 2021. [Google Scholar] [Crossref]
19.
N. A. Taha, T. Hassan, and M. A. Younus, “Face spoofing detection using deep CNN,” Turk. J. Comput. Math. Educ., vol. 12, no. 13, pp. 4363–4373, 2021. [Google Scholar]
20.
S. Shekhar, A. Patel, M. Haloi, and A. Salim, “An ensemble model for face liveness detection,” arXiv preprint, 2022. [Google Scholar] [Crossref]
21.
Z. T. Yu, C. X. Zhao, K. H. M. Cheng, X. M. Cheng, and G. Y. Zhao, “Flexible-modal face anti-spoofing: A benchmark,” arXiv preprint, 2022. [Google Scholar] [Crossref]
22.
Y. Bian, P. Zhang, J. J. Wang, C. M. Wang, and S. L. Pu, “Learning multiple explainable and generalizable cues for face anti-spoofing,” in ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Singapore, Singapore, May 23-27, 2022, pp. 2310–2314. [Google Scholar] [Crossref]
23.
Y. K. Sharma, S. P. Patil, and R. D. Patil, “Deep transfer learning for face spoofing detection,” IOSR J. Comput. Eng., vol. 22, no. 5, pp. 16–20, 2020. [Google Scholar] [Crossref]
24.
J. Z. Guo, X. Y. Zhu, J. C. Xiao, Z. Lei, G. X. Wan, and S. Z. Li, “Improving face anti-spoofing by 3D virtual synthesis,” in 2019 International Conference on Biometrics (ICB), Crete, Greece, February 10, 2019, pp. 1–8. [Google Scholar] [Crossref]
25.
Z. H. Ming, Z. T. Yu, M. Al-Ghadi, M. Visani, M. M. Luqman, and J. C. Burie, “Vitranspad: Video transformer using convolution and self-attention for face presentation attack detection,” in 2022 IEEE International Conference on Image Processing (ICIP), Bordeaux, France, 2022, pp. 4248–4252. [Google Scholar] [Crossref]
26.
Y. H. Zhang, Y. C. Wu, Z. F. Yin, J. Shao, and Z. W. Liu, “Robust face anti-spoofing with dual probabilistic modelling,” arXiv preprint, 2022. [Google Scholar] [Crossref]
27.
S. L. Yang, W. Wang, C. Y. Xu, B. Peng, and J. Dong, “Exposing fine-grained adversarial vulnerability of face anti-spoofing models,” arXiv preprint, 2023. [Google Scholar] [Crossref]
28.
N. Sergievskiy, R. Vlasov, and R. Trusov, “Generalizable method for face anti-spoofing with semi-supervised learning,” arXiv preprint, 2022. [Google Scholar] [Crossref]
29.
S. Chen, T. Xu, Z. Feng, X. Wu, and J. Kittler, “Face anti-spoofing with local difference network and binary facial mask supervision,” J. Electron. Imaging, vol. 31, no. 1, pp. 1–18, 2022. [Google Scholar] [Crossref]
30.
J. Yang, Z. Lei, and S. Z. Li, “Learn convolutional neural network for face anti-spoofing,” arXiv preprint, 2014. [Google Scholar] [Crossref]
31.
M. P. Lai, M. M. Li, and A. R. U. Yasar, “Deep learning for face anti-spoofing: An End-to-End approach,” 2017 Signal Processing: Algorithms, Architectures, Arrangements, and Applications (SPA). pp. 195–200, 2017. [Google Scholar] [Crossref]
32.
D. A. Muhtasim, M. I. Pavel, and S. Y. Tan, “A patch-based CNN built on the VGG-16 architecture for real-time facial liveness detection,” Sustainability, vol. 14, no. 16, pp. 1–11, 2022. [Google Scholar] [Crossref]
Search
Open Access
Research article

Enhancing Face Spoofing Attack Detection: Performance Evaluation of a VGG-19 CNN Model

thomas ayanwola*,
awodele oludele,
michael agbaje
Department of Computer Science, Babcock University, 121003 Ilishan, Nigeria
Acadlore Transactions on AI and Machine Learning
|
Volume 2, Issue 2, 2023
|
Pages 84-98
Received: 04-13-2023,
Revised: 05-09-2023,
Accepted: 05-28-2023,
Available online: 06-12-2023
View Full Article|Download PDF

Abstract:

With the wide use of facial verification and authentication systems, the performance evaluation of Spoofing Attack Detection (SAD) module in the systems is important, because poor performance leads to successful face spoofing attacks. Previous studies on face SAD used a pretrained Visual Geometry Group (VGG) -16 architecture to extract feature maps from face images using the convolutional layers, and trained a face SAD model to classify real and fake face images, obtaining poor performance for unseen face images. Therefore, this study aimed to evaluate the performance of VGG-19 face SAD model. Experimental approach was used to build the model. VGG-19 algorithm was used to extract Red Green Blue (RGB) and deep neural network features from the face datasets. Evaluation results showed that the performance of the VGG-19 face SAD model improved by 6% compared with the state-of-the-art approaches, with the lowest equal error rate (EER) of 0.4%. In addition, the model had strong generalization ability in top-1 accuracy, threshold operation, quality test, fake face test, equal error rate, and overall test standard evaluation metrics.

Keywords: Security, Biometric, Face, Spoofing, VGG-19, Evaluation, Performance

1. Introduction

Biometric user verification and authentication systems have become increasingly popular for applications such as device unlocking, automatic e-transactions, border security, airport control, attendance systems at colleges and universities, face payment, and electronic polling [1]. Among various biometric traits, facial recognition algorithms have gained widespread adoption due to their convenience and contactless nature [2]. These systems work by comparing face images captured by cameras against a database of known face images to identify matches [3].

However, face spoofing attacks (FSAs) pose a significant challenge to the security of facial verification and authentication systems. In FSAs, unauthorized individuals impersonate registered users by obtaining their facial data through social media networks or other means, and then use the acquired data to deceive the system and gain unauthorized access [4], [5]. To ensure robust security in real-world scenarios, face spoofing attack detection (SAD) models must be reliable and demonstrate strong performance on unseen face images [6].

Previous approaches to face SAD mainly relied on extracting Local Binary Patterns (LBP) feature histograms from face patches, such as eyes, nose, and mouth regions, using convolutional layers of fine-tuned Visual Geometry Group (VGG) 11-16 architectures [7]. Despite training various facial parts and three Convolutional Neural Network (CNN) architectures, these models exhibited poor performance and a high rate of successful FSAs [8]. Other studies proposed the use of Discrete Cosine Transform (DCT) and LBP features [9], [10], Histogram of Oriented Gradients (HOG) features [11], and Gabor wavelet features [12]. However, these features lack sufficient distinguishing traits to accurately classify real and fake images, resulting in overfitted face SAD models with poor performance [13].

Existing face SAD models suffer from inadequate training datasets and limited feature extraction capabilities, leading to poor detection of unseen FSAs in real-time scenarios [14], [15]. Although some researchers have explored the application of CNNs for face SAD [16], the potential of the Visual Geometry Group-19 (VGG-19) CNN architecture and joint learning of Red Green Blue (RGB) and deep network features for improved classification has not been fully investigated.

This study aims to evaluate the performance of a VGG-19 face SAD model that leverages joint learning of RGB and deep features to enhance the detection of spoofing attacks. The evaluation results and comparisons with state-of-the-art approaches will be presented in detail, along with an analysis of the model's generalization capabilities across various evaluation metrics.

2. Methodology

The face SAD models, namely, VGG-19A, VGG-19B, VGG-19C, and VGG-19D were implemented using Python programming language, Google Colaboratory, and TensorFlow 2.0 framework [17]. TensorBoard, a visualization tool provided with TensorFlow, was used for tracking loss and accuracy experiment metrics, and visualizing the model graph [18]. Face datasets were used for the end-to-end training of VGG-19 architecture and its three derived network [19]. The primary sources of the face datasets were Nanjing University of Aeronautics and Astronautics (NUAA), Chinese Academy of Sciences' Institute of Automation (CASIA), OUL University (OULU), Wide Multi Channel presentation Attack (WMCA), 3D Mask Attack Dataset (3DMAD), and CASIA-Face-Africa [20]. Then the trained models were tested with unknown face datasets, and evaluated using the metrics in the literature, such as top-1 accuracy, threshold operation accuracy, quality test, fake face test, EER, and overall test [21]. The model with the best evaluation result was used as the face SAD model. Each dataset contained training, validation, and testing face images, which were further divided into real and fake faces [22]. The spoof attacks in the datasets included mask, photo, and video attacks [22]. Table 1 depicts the basic characteristics of each of the six datasets.

The raw face datasets were pre-processed for detecting, cropping and removing dirty face images. Singular Value Decomposition (SVD) tool was used for pre-processing, which detected faces and facial landmarks on images and resized the face images to 244x244x3. The datasets contained 80,000 training set and 16,000 test set after pre-processing [23].

The datasets had totally 85,571 real and fake faces, which were pre-processed into their bit values, and labeled according to RGB pixel using SVD. After removing 5,571 dirty face images, the rest 80,000 real and fake face images remained. 80% of the datasets was the training set and 20% was the test set. The training set had 64,000 face images, with 32,000 real ones and 32,000 fake ones. A subset of training set was used as validation set, with 3,200 real face images and 3,200 fake ones. The test set had 16,000 face images, with 8,000 real ones and 8,000 fake ones.

The input raw face images were pre-processed and visualized before training [24]. The VGG-19A, VGG-19B, VGG-19C, and VGG-19D were trained using RMSProp gradient based optimization algorithm, with an initial learning rate of 0.00002 [25]. The neural networks were trained for a total of 89 to 100 epochs [26]. The TensorFlow optimal learning rate finder was used after 10 and 15 epochs [27].

Considering the size of the face images database, the total time for training model in Google Collaboratory Cloud environment with NVIDIA GeForce RTX 4090 GPU was 16 hours [28]. The face images were extracted from the six datasets to form the face spoofing detection database [29].

Table 1. Basic characteristics of face SAD datasets

S/N

Datasets

Number of subjects

Number of face images

Number of real images

Number of fake images

Modal types

Spoof attack types

1

3DMAD

12

255

100

155

RGB/Depth

Mask

2

CASIA

1000

21000

10500

10500

RGB/Depth/IR

Photo, and video

3

NUAA

15

12614

5105

7509

RGB

Photo

4

OULU

55

5940

1980

3960

RGB

Photo, and video

5

WMCA

72

6716

3358

3358

RGB/Depth/IR/Thermal

Photo, video, and mask

6

CASIA-Face-Africa

1183

38546

19273

19273

RGB

Photo

3. Results

The training and validation results of the VGG-19A model are shown in Table 2 and Figure 1.

Table 2. Training and validation results of the VGG-19A model

Epoch

Loss

Accuracy

Validation loss

Validation accuracy

1/100

0.2858

0.8971

0.5654

0.7672

2/100

0.2240

0.9141

0.5211

0.7988

3/100

0.2058

0.9194

0.4884

0.8100

4/100

0.1944

0.9236

0.5188

0.8064

5/100

0.1868

0.9249

0.5525

0.8050

6/100

0.1812

0.9281

0.6030

0.8016

7/100

0.1741

0.9316

0.5959

0.8084

8/100

0.1702

0.9330

0.5497

0.8140

9/100

0.1656

0.9340

0.6051

0.8100

10/100

0.1609

0.9367

0.5918

0.8118

11/100

0.1565

0.9384

0.6218

0.8104

12/100

0.1520

0.9410

0.5910

0.8182

13/100

0.1477

0.9426

0.5864

0.8152

14/100

0.1443

0.9444

0.5927

0.8190

15/100

0.1399

0.9467

0.6234

0.8182

16/100

0.1365

0.9481

0.6289

0.8176

17/100

0.1317

0.9508

0.5956

0.8240

18/100

0.1285

0.9529

0.6696

0.8190

19/100

0.1258

0.9536

0.6419

0.8246

20/100

0.1218

0.9556

0.7354

0.8206

21/100

0.1173

0.9558

0.6038

0.8296

22/100

0.1138

0.9597

0.8082

0.8110

23/100

0.1100

0.9614

0.6923

0.8226

24/100

0.1065

0.9622

0.8417

0.8096

25/100

0.1031

0.9636

0.8782

0.8106

26/100

0.0995

0.9655

0.7675

0.8228

27/100

0.0965

0.9674

0.7405

0.8272

28/100

0.0924

0.9698

0.7064

0.8308

29/100

0.0898

0.9706

0.8736

0.8150

30/100

0.0849

0.9725

0.8803

0.8116

31/100

0.0822

0.9739

0.8691

0.8182

32/100

0.0788

0.9764

0.8833

0.8130

33/100

0.0758

0.9772

0.8148

0.8254

34/100

0.0723

0.977

0.8831

0.8248

35/100

0.0692

0.979

0.9101

0.8150

36/100

0.0662

0.9809

0.9683

0.8188

37/100

0.0622

0.9826

0.8967

0.8262

38/100

0.0598

0.9839

1.0424

0.8156

39/100

0.0581

0.9844

0.971

0.8218

40/100

0.0537

0.9859

1.1659

0.8110

41/100

0.0510

0.9874

1.0033

0.8192

42/100

0.0489

0.9879

1.0974

0.8192

43/100

0.0457

0.9888

0.9876

0.8292

44/100

0.0435

0.9902

1.0909

0.8204

45/100

0.0401

0.9915

0.9841

0.8322

46/100

0.0385

0.9914

1.1322

0.8232

47/100

0.0361

0.9918

1.1609

0.8202

48/100

0.0339

0.9931

1.1549

0.8208

49/100

0.0319

0.9929

1.1243

0.8284

50/100

0.0296

0.9941

1.1067

0.8276

51/100

0.0265

0.9949

1.4574

0.8112

52/100

0.0253

0.9948

1.2037

0.8274

53/100

0.0237

0.9959

1.3284

0.8202

54/100

0.0215

0.9961

1.2718

0.8276

55/100

0.0199

0.9966

1.3732

0.8202

56/100

0.0187

0.9972

1.4028

0.8210

57/100

0.0166

0.9979

1.2637

0.8262

58/100

0.0152

0.9981

1.3402

0.8216

59/100

0.0139

0.9981

1.4616

0.8240

60/100

0.0128

0.9988

1.5826

0.8158

61/100

0.0115

0.9986

1.4762

0.8256

62/100

0.0107

0.9990

1.3859

0.8314

63/100

0.0096

0.9989

1.5240

0.8270

64/100

0.0086

0.9991

1.4305

0.8292

65/100

0.0078

0.9995

1.6003

0.8260

66/100

0.0067

0.9997

1.7436

0.8234

67/100

0.0061

0.9994

1.5660

0.8238

68/100

0.0053

0.9995

1.7274

0.8258

69/100

0.0050

0.9996

1.6954

0.8240

70/100

0.0042

0.9997

1.6457

0.8290

71/100

0.0036

0.9997

1.9800

0.8188

72/100

0.0033

0.9997

1.9834

0.8156

73/100

0.0028

0.9998

1.7679

0.8282

74/100

0.0024

0.9997

1.9357

0.8208

75/100

0.0022

0.9999

2.0402

0.8268

76/100

0.0016

0.9999

2.1737

0.8236

77/100

0.0017

0.9998

2.1998

0.8240

78/100

0.0013

0.9999

1.9325

0.8294

79/100

0.0012

0.9999

2.2198

0.8266

80/100

0.0011

0.9998

2.0579

0.8252

81/100

8.35E-04

0.9999

2.2029

0.8280

82/100

8.06E-04

0.9998

2.1431

0.8300

83/100

4.82E-04

0.9999

2.1699

0.8304

84/100

3.99E-04

1

2.5538

0.8244

85/100

4.21E-04

1

2.3126

0.8284

86/100

5.41E-04

0.9999

2.2783

0.8310

87/100

2.51E-04

1

2.5194

0.8276

88/100

2.86E-04

1

2.6917

0.8232

89/100

2.51E-04

1

2.6306

0.8268

Figure 1. Training and validation accuracy of the VGG-19A model

The training and validation results of the VGG-19B model are shown in Table 3 and Figure 2.

Table 3. Training and validation results of the VGG-19B model

Epoch

Loss

Accuracy

Validation loss

Validation accuracy

1/100

0.3138

0.8882

0.5966

0.7788

2/100

0.244

0.9079

0.4973

0.8082

3/100

0.2229

0.914

0.5208

0.8082

4/100

0.2122

0.918

0.5499

0.8086

5/100

0.2029

0.9195

0.5629

0.8056

6/100

0.1977

0.9221

0.5736

0.8054

7/100

0.1911

0.9245

0.6038

0.8018

8/100

0.185

0.9272

0.5207

0.8186

9/100

0.1817

0.927

0.5646

0.8154

10/100

0.1788

0.9301

0.612

0.8150

11/100

0.1749

0.9316

0.6354

0.8132

12/100

0.1731

0.9317

0.5743

0.8200

13/100

0.1676

0.9329

0.5841

0.8196

14/100

0.1641

0.9381

0.6317

0.8176

15/100

0.1628

0.9361

0.6492

0.8176

16/100

0.1583

0.9403

0.6632

0.8188

17/100

0.1554

0.9419

0.5346

0.8294

18/100

0.1523

0.9413

0.7164

0.8118

19/100

0.1491

0.943

0.6535

0.8206

20/100

0.1468

0.9442

0.7222

0.8170

21/100

0.1436

0.944

0.6846

0.8158

22/100

0.141

0.9451

0.6906

0.8218

23/100

0.1384

0.9456

0.735

0.8198

24/100

0.1318

0.9508

0.7863

0.8146

25/100

0.1329

0.9523

0.7608

0.8206

26/100

0.1294

0.9523

0.7852

0.8210

27/100

0.1256

0.9555

0.7489

0.8228

28/100

0.1241

0.9547

0.7982

0.8166

29/100

0.1197

0.9565

0.7759

0.8250

30/100

0.1159

0.9587

0.7712

0.8218

31/100

0.1154

0.9584

0.6788

0.8320

32/100

0.1108

0.9618

0.78

0.8270

33/100

0.1077

0.9624

0.7895

0.8286

34/100

0.1052

0.9644

0.7135

0.8322

35/100

0.1022

0.9644

0.7716

0.8274

36/100

0.0979

0.9686

0.7806

0.8300

37/100

0.0963

0.9678

0.9526

0.8140

38/100

0.0949

0.969

0.8005

0.8284

39/100

0.0917

0.9689

0.822

0.8296

40/100

0.0865

0.9729

0.8628

0.8304

41/100

0.0823

0.9738

0.8923

0.8286

42/100

0.0812

0.9755

0.9835

0.8226

43/100

0.0796

0.9775

0.9474

0.8254

44/100

0.0757

0.9771

1.0239

0.8212

45/100

0.0731

0.9792

0.9554

0.8292

46/100

0.0689

0.9804

1.0073

0.8230

47/100

0.0682

0.9809

1.0359

0.8228

48/100

0.0651

0.9827

1.061

0.8242

49/100

0.0615

0.9843

0.9835

0.8276

50/100

0.0598

0.9836

1.0075

0.8288

51/100

0.0583

0.9859

0.977

0.8296

52/100

0.0542

0.9857

1.2336

0.8098

53/100

0.0519

0.9882

0.9897

0.8340

54/100

0.0492

0.9881

1.223

0.8200

55/100

0.0468

0.9891

1.0489

0.8264

56/100

0.0456

0.9905

1.2621

0.8132

57/100

0.0434

0.9901

1.1457

0.8260

58/100

0.0396

0.9907

1.0934

0.8302

59/100

0.0395

0.9911

1.1126

0.8340

60/100

0.0365

0.9918

1.3168

0.8164

61/100

0.0354

0.9921

1.2735

0.8246

62/100

0.032

0.9936

1.432

0.8196

63/100

0.0308

0.9937

1.2756

0.8330

64/100

0.0297

0.9941

1.179

0.8366

65/100

0.0272

0.9946

1.3122

0.8308

66/100

0.0253

0.9952

1.4634

0.8196

67/100

0.0247

0.9955

1.4153

0.8226

68/100

0.0233

0.9961

1.2986

0.8330

69/100

0.0211

0.996

1.3651

0.8308

70/100

0.0215

0.9967

1.4366

0.8276

71/100

0.0185

0.9968

1.3978

0.8314

72/100

0.0170

0.9973

1.6028

0.8226

73/100

0.0163

0.9971

1.3698

0.8302

74/100

0.0154

0.9976

1.8460

0.8126

75/100

0.0142

0.9975

1.5969

0.8268

76/100

0.0135

0.9979

1.6746

0.8228

77/100

0.0120

0.9981

1.5466

0.8332

78/100

0.0114

0.9983

1.6801

0.8258

79/100

0.0105

0.9984

1.5240

0.8332

80/100

0.0094

0.9990

1.7059

0.8266

81/100

0.0088

0.9989

2.1167

0.8044

82/100

0.0076

0.9991

1.6928

0.8266

83/100

0.0073

0.9993

1.7823

0.8276

84/100

0.0060

0.9992

1.8105

0.8294

85/100

0.0061

0.9994

2.0048

0.8260

86/100

0.0058

0.9993

2.0594

0.8232

87/100

0.0047

0.9994

2.0616

0.8224

88/100

0.0041

0.9995

2.0568

0.8236

89/100

0.0038

0.9997

2.1799

0.8214

90/100

0.0034

0.9997

2.2058

0.8304

91/100

0.0030

0.9999

2.3742

0.8430

92/100

0.0027

0.9999

1.8925

0.8674

93/100

0.0025

0.9996

2.2388

0.8762

94/100

0.0024

0.9997

2.4344

0.8886

95/100

0.0020

0.9998

2.4973

0.8986

96/100

0.0019

0.9999

1.8693

0.9014

97/100

0.0014

0.9999

2.3205

0.9284

98/100

0.0011

1

2.3119

0.9394

99/100

9.97E-04

0.9999

2.6476

0.9510

100/100

8.95E-04

1

2.4416

0.9755

Figure 2. Training and validation accuracy of the VGG-19B model

The training and validation results of the VGG-19C model are shown in Table 4 and Figure 3.

Table 4. Training and validation results of the VGG-19C model

Epoch

Loss

Accuracy

Validation loss

Validation accuracy

1/100

0.3349

0.8756

0.5688

0.7760

2/100

0.2538

0.9036

0.5303

0.7902

3/100

0.2343

0.9118

0.5928

0.7926

4/100

0.2216

0.9153

0.5179

0.8126

5/100

0.2112

0.9194

0.5443

0.8152

6/100

0.2041

0.9201

0.5973

0.8078

7/100

0.1989

0.9216

0.6048

0.8114

8/100

0.1954

0.9244

0.5916

0.8106

9/100

0.1934

0.9232

0.6603

0.8072

10/100

0.1882

0.9272

0.6880

0.8078

11/100

0.1861

0.9276

0.7364

0.8040

12/100

0.1840

0.9285

0.7216

0.8048

13/100

0.1820

0.9293

0.7575

0.8028

14/100

0.1787

0.9296

0.7229

0.8082

15/100

0.1755

0.9314

0.6675

0.8088

16/100

0.1723

0.9315

0.7160

0.8098

17/100

0.1713

0.9352

0.6333

0.8176

18/100

0.1685

0.9366

0.7949

0.8054

19/100

0.1634

0.9395

0.6953

0.8166

20/100

0.164

0.9361

0.7947

0.8088

21/100

0.1606

0.9391

0.7571

0.8132

22/100

0.1567

0.9400

0.8629

0.8062

23/100

0.1534

0.9420

0.8066

0.8112

24/100

0.1525

0.9412

0.8630

0.8084

25/100

0.1501

0.9441

0.9439

0.8062

26/100

0.1476

0.9451

0.8715

0.8088

27/100

0.1458

0.9442

0.9569

0.8060

28/100

0.1431

0.9485

0.9511

0.8046

29/100

0.1405

0.9480

0.8984

0.8132

30/100

0.1340

0.9499

0.9540

0.8108

31/100

0.1356

0.9501

0.7550

0.8214

32/100

0.1306

0.9545

0.8648

0.8160

33/100

0.1298

0.9534

0.8689

0.8154

34/100

0.1268

0.9569

0.8163

0.8228

35/100

0.1240

0.9557

0.8959

0.8158

36/100

0.1236

0.9577

1.0246

0.8094

37/100

0.1190

0.9587

0.9895

0.8158

38/100

0.1172

0.9602

1.1166

0.8060

39/100

0.1120

0.9610

1.0929

0.8104

40/100

0.1097

0.9627

0.9209

0.8230

41/100

0.1079

0.9659

1.1096

0.8064

42/100

0.1034

0.9657

1.0814

0.8146

43/100

0.0999

0.9679

1.1010

0.8168

44/100

0.0974

0.9691

1.0124

0.8216

45/100

0.0970

0.9700

1.1169

0.8188

46/100

0.0944

0.9693

1.1937

0.8122

47/100

0.0888

0.9709

1.0374

0.8246

48/100

0.089

0.9709

1.0962

0.8198

49/100

0.0861

0.9733

1.0966

0.8228

50/100

0.0841

0.9741

0.9950

0.8276

51/100

0.0791

0.9769

1.1169

0.8194

52/100

0.0781

0.9764

1.1598

0.8220

53/100

0.0740

0.9795

1.2084

0.8198

54/100

0.0732

0.9800

1.202

0.8206

55/100

0.0696

0.9814

1.3061

0.8146

56/100

0.0659

0.9821

1.306

0.8176

57/100

0.0653

0.9816

1.3766

0.8148

58/100

0.0636

0.9825

1.4566

0.8102

59/100

0.0592

0.9851

1.3576

0.8182

60/100

0.0573

0.9856

1.3334

0.8216

61/100

0.0557

0.9868

1.4545

0.8126

62/100

0.0523

0.9859

1.3008

0.8260

63/100

0.0523

0.9883

1.4315

0.8186

64/100

0.0481

0.9876

1.2682

0.8300

65/100

0.0471

0.9884

1.4398

0.8244

66/100

0.0427

0.9901

1.5312

0.8148

67/100

0.0392

0.9908

1.5541

0.8224

68/100

0.0396

0.9903

1.5583

0.8220

69/100

0.0381

0.9915

1.5567

0.8200

70/100

0.0351

0.9927

1.693

0.8142

71/100

0.0347

0.9921

1.6603

0.8176

72/100

0.0331

0.9934

1.4412

0.8264

73/100

0.0313

0.9934

1.6319

0.8232

74/100

0.0280

0.9948

1.5561

0.8244

75/100

0.0258

0.9948

1.6204

0.8244

76/100

0.0250

0.9951

1.7106

0.8198

77/100

0.0255

0.9948

1.7369

0.8248

78/100

0.0226

0.9949

1.707

0.8250

79/100

0.0204

0.9966

1.7334

0.8262

80/100

0.0183

0.9969

1.9623

0.8112

81/100

0.0188

0.9973

1.7961

0.8242

82/100

0.0169

0.9967

1.7671

0.8254

83/100

0.0172

0.9970

1.7826

0.8264

84/100

0.0151

0.9973

1.9336

0.8216

85/100

0.0126

0.9982

2.1171

0.8186

86/100

0.0128

0.9980

1.9428

0.8228

87/100

0.0120

0.9984

2.2851

0.8246

88/100

0.0102

0.9988

1.9325

0.8346

89/100

0.0105

0.9981

1.901

0.8468

90/100

0.0097

0.9991

2.0587

0.8564

91/100

0.0088

0.9990

2.1058

0.8642

92/100

0.0089

0.9987

2.3508

0.8760

93/100

0.0075

0.9988

1.9600

0.8812

94/100

0.0070

0.9993

2.0539

0.8960

95/100

0.0067

0.9992

2.2851

0.9116

96/100

0.0058

0.9996

2.2477

0.9250

97/100

0.0051

0.9993

2.3583

0.9308

98/100

0.0057

0.9991

2.5974

0.9462

99/100

0.0048

0.9997

2.4989

0.9752

100/100

0.0043

0.9996

2.1695

0.9883

Figure 3. Training and validation accuracy of the VGG-19C model

The training and validation results of the VGG-19D model are shown in Table 5 and Figure 4.

Table 5. Summary of training and validation results

Epoch

Loss

Accuracy

Validation loss

Validation accuracy

1/100

0.3108

0.8888

0.5443

0.7724

2/100

0.2398

0.9063

0.4871

0.8112

3/100

0.2203

0.9149

0.5840

0.7914

4/100

0.2084

0.9182

0.5894

0.7964

5/100

0.2002

0.9203

0.5687

0.8040

6/100

0.1953

0.9249

0.5676

0.8074

7/100

0.1885

0.9250

0.5618

0.8062

8/100

0.1839

0.9257

0.5976

0.8058

9/100

0.1781

0.9295

0.6025

0.8054

10/100

0.1766

0.9300

0.6301

0.8068

11/100

0.1726

0.9312

0.5985

0.8092

12/100

0.1692

0.9344

0.6517

0.8068

13/100

0.1655

0.9364

0.6528

0.8068

14/100

0.1606

0.9389

0.6385

0.8102

15/100

0.1588

0.9386

0.6722

0.8078

16/100

0.1538

0.9398

0.6290

0.8128

17/100

0.1505

0.9449

0.6300

0.8150

18/100

0.1481

0.9435

0.6189

0.8150

19/100

0.1457

0.9453

0.7476

0.8086

20/100

0.1419

0.9468

0.6346

0.8222

21/100

0.1403

0.9484

0.7247

0.8140

22/100

0.1334

0.9491

0.7873

0.8100

23/100

0.1313

0.9524

0.7193

0.8142

24/100

0.1270

0.9545

0.7466

0.8156

25/100

0.1261

0.9531

0.7458

0.8184

26/100

0.1214

0.9561

0.8872

0.8058

27/100

0.1185

0.959

0.6927

0.8198

28/100

0.1162

0.9588

0.9095

0.8076

29/100

0.1141

0.9601

0.6847

0.8260

30/100

0.1090

0.9606

0.8041

0.8160

31/100

0.1072

0.9606

1.0424

0.8050

32/100

0.1035

0.9656

0.9668

0.8104

33/100

0.1002

0.9652

0.8691

0.8146

34/100

0.0981

0.9669

0.788

0.8232

35/100

0.0946

0.9694

0.7872

0.8218

36/100

0.0907

0.9693

1.0007

0.8102

37/100

0.0880

0.9713

0.897

0.8158

38/100

0.0838

0.973

0.8293

0.8276

39/100

0.0800

0.9764

0.888

0.8230

40/100

0.0771

0.9773

0.9242

0.8164

41/100

0.0741

0.9785

1.1113

0.8132

42/100

0.0719

0.9789

0.9078

0.8232

43/100

0.0693

0.9801

1.0677

0.8160

44/100

0.0654

0.9818

1.0153

0.8192

45/100

0.0624

0.9828

0.909

0.8278

46/100

0.0599

0.9845

1.0636

0.8196

47/100

0.0575

0.9839

1.4296

0.7994

48/100

0.0553

0.9858

1.0433

0.8234

49/100

0.0529

0.9871

1.2953

0.8094

50/100

0.0483

0.9876

1.1795

0.8202

51/100

0.0474

0.9895

1.0696

0.8226

52/100

0.0447

0.9893

1.1296

0.8216

53/100

0.0419

0.9894

1.1918

0.8224

54/100

0.0399

0.9905

1.2458

0.8212

55/100

0.0368

0.9929

1.2821

0.8192

56/100

0.0353

0.9924

1.2305

0.8244

57/100

0.0332

0.993

1.5264

0.8074

58/100

0.0325

0.9939

1.0668

0.8340

59/100

0.0298

0.9947

1.2203

0.8252

60/100

0.0276

0.9943

1.3597

0.8240

61/100

0.0263

0.9955

1.2439

0.8278

62/100

0.0249

0.9951

1.3958

0.8222

63/100

0.0231

0.9956

1.4306

0.8204

64/100

0.0223

0.9959

1.3825

0.8258

65/100

0.0203

0.996

1.601

0.8152

66/100

0.0188

0.997

1.4347

0.8230

67/100

0.0176

0.9975

1.4804

0.8256

68/100

0.0160

0.9978

1.4836

0.8252

69/100

0.0152

0.9976

1.5674

0.8208

70/100

0.0140

0.9981

1.5481

0.8232

71/100

0.0130

0.9984

1.5227

0.8214

72/100

0.0121

0.9983

1.417

0.8314

73/100

0.0111

0.9989

1.4488

0.8360

74/100

0.0101

0.9985

1.4712

0.8300

75/100

0.0093

0.9988

1.5764

0.8264

76/100

0.0084

0.9989

1.6357

0.8268

77/100

0.0074

0.9989

1.8639

0.8194

78/100

0.0077

0.9991

1.7494

0.8242

79/100

0.0064

0.9993

1.9100

0.8218

80/100

0.0058

0.9993

1.9576

0.8158

81/100

0.0056

0.9993

2.0104

0.8204

82/100

0.0042

0.9996

1.9334

0.8230

83/100

0.0043

0.9996

1.9391

0.8234

84/100

0.0040

0.999

2.0536

0.8198

85/100

0.0031

0.9997

1.9390

0.8280

86/100

0.0024

0.9997

1.8474

0.8332

87/100

0.0023

0.9998

1.9956

0.8246

88/100

0.0020

0.9999

2.0974

0.8232

89/100

0.0020

0.9997

2.1888

0.8224

90/100

0.0013

0.9999

2.0936

0.8272

91/100

0.0014

0.9999

2.1994

0.8358

92/100

0.0014

0.9998

1.9263

0.8440

93/100

0.0010

1

2.3412

0.8758

94/100

0.0009

0.9999

2.3328

0.8952

95/100

0.0007

0.9999

2.5126

0.9046

96/100

0.0006

1

2.5790

0.9248

Figure 4. Training and validation accuracy of the VGG-19D model

The summary of both training and validation results and test results of the four models is shown in Table 6 and Table 7, respectively.

Table 6. Summary of training and validation results

Models

Loss

Accuracy

Validation loss

Validation accuracy

VGG-19A

2.14E-05

100

2.98

97.02

VGG-19B

8.95E-04

100

2.45

97.55

VGG-19C

0.0043

99.96

1.17

98.83

VGG-19D

0.00033169

100

2.47

97.53

Table 7. Summary of test results

Networks

Accuracy

Loss

VGG-19A

97.65

2.35

VGG-19B

98.13

1.87

VGG-19C

99.16

0.84

VGG-19D

98.95

1.05

Performance of VGG-19A, VGG-19B, VGG-19C, and VGG-19D models was evaluated using several standard performance evaluation metrics, such as top-1 accuracy, threshold operation, quality test, fake face test, and overall test. The four VGG-19 algorithms were trained and tested on the six datasets. The datasets had 2,337 real subjects, and their fake faces were generated using wrapped photo, cut-photo, mask and video attacks, respectively. The low, normal, and high quality imaging was considered. The datasets provided 80,000 face images, with 40,000 real faces and 40,000 fake ones.

The top-1 accuracy measured the proportion of face images with the predicted label matching the single target label. The output class with the highest probability was considered. The results of class output probability as well as real and fake face output are depicted in Table 8 and Table 9.

Table 8. Probability of class output

Positive

Negative

True

1

1

False

0

0

Table 9. Real and fake face output

Input

Output

Real face

Real face (true positive)

Fake face (false negative)

Fake face

Fake face (true negative)

Real face (false positive)

The top-1 accuracy result of VGG-19A, VGG-19B, VGG-19C, and VGG-19D models is depicted in Table 10.

It can be seen from Table 10 that the VGG-19C obtained an average top-1 accuracy of 99%. Dropout rate of 0.5 was utilized for additional regularization after every max-pooling layer in VGG-19C, which reduced the amount of overfitting in the model, thus causing a strong generalization ability.

Threshold operation accuracy considered only the probability of real face class. The probability of the true class obtained in the output of the VGG-19 models was compared with a set of threshold values to determine whether the output probability was higher than the corresponding threshold. It can be observed from Table 11 that VGG-19C has achieved the best performance. Compared with the state-of-the-art approaches [30], [31], [32], VGG-19C achieved an average test accuracy of 99% with the lowest EER of 0.4%.

Table 10. Top-1 accuracy

SAD models

Top-1 accuracy

VGG-19A

97%

VGG-19B

98%

VGG-19C

99%

VGG-19D

98%

Table 11. Threshold operation accuracy

SAD models

Threshold operation accuracy

EER

VGG-19A

$85 \%$

$7 \%$

VGG-19B

$82 \%$

$7 \%$

VGG-19C

$99 \%$

$0.4 \%$

VGG-19D

$90 \%$

$8 \%$

CNN + SVM [30]

$81 \%$

$7.49 \%$

VGG-11 [31]

$89 \%$

$5 \%$

VGG-16 [32]

$97 \%$

$0.67 \%$

This quality test was used to evaluate the performance of the SAD model given that the input image quality was fixed. The images of each subject were captured concurrently using three cameras with low, normal, and high quality. The SVD face detector was used to classify the image quality, and the quality test result on the datasets is depicted in Table 12.

Table 12. Quality test accuracy
SAD modelsNormalLowHigh
VGG-19A$91 \%$$84 \%$$78 \%$
VGG-19B$94 \%$$86 \%$$80 \%$
VGG-19C$97 \%$$94 \%$$84 \%$
VGG-19D$95 \%$$87 \%$$82 \%$

The fake face test was used to evaluate the performance of the face SAD model given that the fake face image types were fixed. The fake face images were in wrapped photo, cut photo, video, and mask attacks. The significance of this metric was to have as many as possible fake face image samples to train and test the VGG-19 model. The result is depicted in Table 13.

Table 13. Fake face test accuracy
SAD modelsWrapped photo attackCut photo attackVideo attackMask attack
VGG-19A$79 \%$$80 \%$$82 \%$$89 \%$
VGG-19B$89 \%$$82 \%$$82 \%$$87 \%$
VGG-19C$98 \%$$87 \%$$89 \%$$88 \%$
VGG-19D$78 \%$$90 \%$$90 \%$$90 \%$

It can be observed from Table 12 that the VGG-19C model has obtained an overall high classification rate, compared with VGG-19A, VGG-19B, and VGG-19D. Similarly, VGG-19C has obtained an overall high accuracy for all four types of attacks in fake face test, as shown in Table 13.

The overall test was used to evaluate the general performance of the face SAD model by combining all the data. Table 14 compares the quality test using the proposed VGG-19 model for the face SAD model with other state-of-the-art approaches using CNN with SVM, VGG-11, and VGG-16 for developing the model.

Table 14. Overall test accuracy

SAD models

Low

Normal

High

VGG-11 [31]

94%

94%

82%

CNN + SVM [30]

75%

83%

90%

VGG-16 [32]

91%

90%

89%

VGG-19C

97%

96%

95%

4. Discussion

The VGG-19 base architecture was designed for different image classification types. In order to achieve the best performance in classifying human face images, VGG-19B, VGG-19C, and VGG-19D were derived from the base architecture and trained on the same normalized extracted features. The multi-ethnicity face datasets across the entire globe were used for the training of the SAD model.

Extraction of RGB and deep network features from the face images produced more distinct traits for the training and validation of the face SAD model. The normalized extracted features were used to train, validate, and test VGG-19A, VGG-19B, VGG-19C, and VGG-19D models. VGG-19A, VGG-19B, and VGG-19D obtained 100% accuracy while VGG-19C achieved 99.96% accuracy. However, VGG-19C was the best SAD model because it had the best validation accuracy of 98.83%. VGG-19C achieved 0.4% EER and an improvement of 6% in overall test compared with the state-of-the-art approaches, after being evaluated using top-1 accuracy, threshold operation accuracy, quality test, fake face test, overall test and benchmark to CNN with SVM, VGG-11, and VGG-16.

5. Conclusions

Face datasets were collected from CASIA, NUAA, OULU, WMCA, 3DMAD, and CASIA-Face-Africa, and pre-processed using SVD. Dropout regularization technique was added to neural network layers to overcome overfitting. VGG-19 architecture was used to extract features from RGB and deep network to classify real and fake face images. The designed neural networks were implemented with TensorFlow 2.0 framework. The extracted features were normalized, and used to train the four VGG-19 neural networks. The trained networks were tested with unknown face datasets, which showed that VGG-19C was the best face SAD model with 99.96% training accuracy, 98.83% validation accuracy, and 99.16% testing accuracy. In addition, VGG-19C was evaluated using top-1 accuracy, threshold operation accuracy, quality test, fake face test, overall test and benchmark to face SAD models of CNN with SVM, VGG-11, and VGG-16, which showed that VGG-19C achieved 0.4% EER, and an improvement of 6% in overall test compared with the state-of-the-art approaches.

Although the study of face SAD has been progressing tremendously using new algorithms and techniques, some open research issues need to be addressed to develop a robust face SAD model for real-time face verification and authentication applications. The most common issue with all face SAD systems is their generalization ability in unconstrained scenes. Apart from face biometric verification and authentication for access control, other potential applications of face SAD need to be explored, such as recognizing live people and photos for self-driving assistance and robot navigation.

Data Availability

The study used CASIA, NUAA, WMCA, OULU, 3DMAD, and CASIA-Face-Africa datasets. The datasets were collected through a research agreement form after which the datasets were released. This procedure was followed to collect the face datasets from owners of the spoofing attack detection datasets:

1. An email was sent for use of the spoofing attacks datasets for academic research purpose using my Babcock student email address.

2. The license agreement form was filled in conjunction with my Thesis supervisors.

3. The filled license agreement form was scanned and sent back.

4. The face datasets were downloaded to my google drive for training, validating and testing of the face SAD models.

5. Extraction of face datasets was done using ZIP Extractor which is compatible with Google Drive.

Acknowledgments

My profound and honest appreciation goes to my research supervisors, Professor O. Awodele; Dr. M. Agbaje; and Dr. A. Ajayi of the School of Computing and Engineering Sciences, Babcock University, Ilishan, for availing me a most conducive environment to carry out this research and for the priceless supervision they provided in the course of this program. I am deeply inspired, motivated, touched, and cared for by their patience, professional comments, dedication, and input. They played key roles that culminated in the accomplishment of this thesis.

I also acknowledge the immense backing of the entire staff of the Department of Computer Science. I thank Professor S. Idowu, Dean School of Computing and Engineering Sciences, Dr. S. Kuroyo, Head of Department Computer Science, Professor S. Okolie, PG School Coordinator, Dr E.E. Onuiri, PG Department Coordinator, Dr S. Maitanmi, School of Sciences Methodologist, Professors A. Ogonna; M. Eze; A. Adebayo, and Drs. O. Ebiesuwa; A. Omotunde, T. Adigun, U. Nzenwatta, C. Ajaegbu, F. Ayankoya, O. Akande, and A. Izang. I thank them for making the Department a warm and conducive environment for learning.

I am particularly indebted to my parents: Mr. Ayanwola Benjamin (late) and Mrs. Mary Ayanwola (nee Oladele) for their love, prayers, care, and sacrifice to ensure that I have access to quality education and a secured future. I am very much thankful to my beautiful wife, Mrs. Dolapo Bosede Ayanwola (nee Agboola), and my lovely God-given children: Ayanwola Faith Oluwanifemi, Ayanwola David Oluwatimilehin, Ayanwola Emmanuel Jesutofunmi, and Ayanwola Joseph Eri-Oluwa for their love, understanding, prayers, and continuing support throughout the program. My special thanks go to my pastor and father in the Lord, Pastor Noruwa Edokpolo, the Provincial Pastor of Lagos Province 77, Redeemed Christian Church of God, for all the opportunities given to me to progress academically.

I am grateful for the administrative assistance, information shared, printing, and editorial support received from some members of the Babcock University community. I particularly acknowledge the support of Mr. Seun Idowu, Mr. Martins Nkume, and Mrs. Jane Chukwuemeka.

Finally, my deepest thanks go to my siblings, entire family, and friends who supported me in one way or the other through the course of this work.

Conflicts of Interest

The authors declare no conflict of interest.

References
1.
B. Choudhury, M. Haldar, B. Issac, V. Raman, and P. Then, “A survey on biometrics and cancelable biometrics systems,” Int. J. Image Graph., vol. 18, no. 1, pp. 1–39, 2018. [Google Scholar] [Crossref]
2.
D. Kalenichenko, J. Philbin, and F. Schroff, “Facenet: A unified embedding for face recognition and clustering,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, June 07-12, 2015. pp. 815–823. [Google Scholar] [Crossref]
3.
C. Busch and R. Ramachandra, “Presentation attack detection methods for face recognition systems: A comprehensive survey,” ACM Comput. Surv., vol. 50, no. 1, pp. 1–37, 2017. [Google Scholar] [Crossref]
4.
R. H. Deng, Y. Li, K. Xu, Y. J. Li, and Q. Yan, “Understanding OSN-based facial disclosure against face authentication systems,” Proceedings of the 9th ACM Symposium on Information, Computer and Communications Security, Kyoto, Japan, June 4-6, 2014. pp. 413–424. [Google Scholar] [Crossref]
5.
X. Y. Tan, Y. Li, J. Liu, and L. Jiang, “Face liveness detection from a single image with sparse low rank bilinear discriminative model,” Computer Vision – ECCV 2010. ECCV 2010. Lecture Notes in Computer Science. Springer, Berlin, Heidelberg, pp. 504–517, 2010. [Google Scholar] [Crossref]
6.
A. Orjiude, “Eight banks lose N1.9bn to fraud in one year,” Punch Newspaper, 2021. https://punchng.com/eight-banks-lose-n1-9bn-to-fraud-in-one-year/ [Google Scholar]
7.
Z. F. Li, Y. Qiao, K. P. Zhang, and Z. P. Zhang, “Joint face detection and alignment using multitask cascaded convolutional networks,” IEEE Signal Process. Lett., vol. 23, no. 10, pp. 1499–1503, 2016. [Google Scholar] [Crossref]
8.
Z. W. Zhang, J. J. Yan, S. F. Liu, Z. Li, Z. Lei, D. Yi, and S. Z. Li, “A face anti-spoofing database with diverse attacks,” in 5th IAPR International Conference on Biometrics (ICB), New Delhi, India, March 29, 2012-April 01, 2012, pp. 26–31. [Google Scholar] [Crossref]
9.
T. Mäenpää, T. Ojala, and M. Pietikäinen, “Multiresolution grayscale and rotation invariant texture classification with local binary patterns,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 7, pp. 971–987, 2002. [Google Scholar] [Crossref]
10.
J. Komulainen, A. Hadid, and M. Pietikäinen, “Face spoofing detection from single images using texture and local shape analysis,” IET Biometrics, vol. 1, no. 1, pp. 3–10, 2012. [Google Scholar] [Crossref]
11.
N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” Proceedings of IEEE Computing Society Conference on Computer Vision and Pattern Recognition, San Diego, CA, USA. pp. 886–893, 2005. [Google Scholar] [Crossref]
12.
B. S. Manjunath and W. Y. Ma, “Texture features for browsing and retrieval of image data,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 18, no. 8, pp. 837–842, 1996. [Google Scholar] [Crossref]
13.
D. Gragnaniello, G. Poggi, C. Sansone, and L. Verdoliva, “An investigation of local descriptors for biometric spoofing detection,” IEEE Trans. Inf. Forensics Secur., vol. 10, no. 4, pp. 849–863, 2015. [Google Scholar] [Crossref]
14.
H. X. Li, Z. Lin, X. H. Shen, J. Brandt, and G. Hua, “A convolutional neural network cascade for face detection,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, June 07-12, 2015. pp. 5325–5334. [Google Scholar] [Crossref]
15.
L. Li, X. Y. Feng, X. Y. Jiang, Z. Q. Xia, and A. Hadid, “Face anti-spoofing via deep local binary patterns,” Proceedings of IEEE International Conference on Image Processing, Beijing, China, September 17-20, 2017. pp. 101–105. [Google Scholar] [Crossref]
16.
K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint, 2014. [Google Scholar] [Crossref]
17.
O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and F. F. Li, “ImageNet large scale visual recognition challenge,” Int. J. Comput. Vis., vol. 115, no. 3, pp. 211–252, 2015. [Google Scholar] [Crossref]
18.
Z. Zhang, C. Jiang, X. Zhong, C. Song, and Y. Zhang, “Two-stream convolutional networks for multi-frame face anti-spoofing,” arXiv preprint, 2021. [Google Scholar] [Crossref]
19.
N. A. Taha, T. Hassan, and M. A. Younus, “Face spoofing detection using deep CNN,” Turk. J. Comput. Math. Educ., vol. 12, no. 13, pp. 4363–4373, 2021. [Google Scholar]
20.
S. Shekhar, A. Patel, M. Haloi, and A. Salim, “An ensemble model for face liveness detection,” arXiv preprint, 2022. [Google Scholar] [Crossref]
21.
Z. T. Yu, C. X. Zhao, K. H. M. Cheng, X. M. Cheng, and G. Y. Zhao, “Flexible-modal face anti-spoofing: A benchmark,” arXiv preprint, 2022. [Google Scholar] [Crossref]
22.
Y. Bian, P. Zhang, J. J. Wang, C. M. Wang, and S. L. Pu, “Learning multiple explainable and generalizable cues for face anti-spoofing,” in ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Singapore, Singapore, May 23-27, 2022, pp. 2310–2314. [Google Scholar] [Crossref]
23.
Y. K. Sharma, S. P. Patil, and R. D. Patil, “Deep transfer learning for face spoofing detection,” IOSR J. Comput. Eng., vol. 22, no. 5, pp. 16–20, 2020. [Google Scholar] [Crossref]
24.
J. Z. Guo, X. Y. Zhu, J. C. Xiao, Z. Lei, G. X. Wan, and S. Z. Li, “Improving face anti-spoofing by 3D virtual synthesis,” in 2019 International Conference on Biometrics (ICB), Crete, Greece, February 10, 2019, pp. 1–8. [Google Scholar] [Crossref]
25.
Z. H. Ming, Z. T. Yu, M. Al-Ghadi, M. Visani, M. M. Luqman, and J. C. Burie, “Vitranspad: Video transformer using convolution and self-attention for face presentation attack detection,” in 2022 IEEE International Conference on Image Processing (ICIP), Bordeaux, France, 2022, pp. 4248–4252. [Google Scholar] [Crossref]
26.
Y. H. Zhang, Y. C. Wu, Z. F. Yin, J. Shao, and Z. W. Liu, “Robust face anti-spoofing with dual probabilistic modelling,” arXiv preprint, 2022. [Google Scholar] [Crossref]
27.
S. L. Yang, W. Wang, C. Y. Xu, B. Peng, and J. Dong, “Exposing fine-grained adversarial vulnerability of face anti-spoofing models,” arXiv preprint, 2023. [Google Scholar] [Crossref]
28.
N. Sergievskiy, R. Vlasov, and R. Trusov, “Generalizable method for face anti-spoofing with semi-supervised learning,” arXiv preprint, 2022. [Google Scholar] [Crossref]
29.
S. Chen, T. Xu, Z. Feng, X. Wu, and J. Kittler, “Face anti-spoofing with local difference network and binary facial mask supervision,” J. Electron. Imaging, vol. 31, no. 1, pp. 1–18, 2022. [Google Scholar] [Crossref]
30.
J. Yang, Z. Lei, and S. Z. Li, “Learn convolutional neural network for face anti-spoofing,” arXiv preprint, 2014. [Google Scholar] [Crossref]
31.
M. P. Lai, M. M. Li, and A. R. U. Yasar, “Deep learning for face anti-spoofing: An End-to-End approach,” 2017 Signal Processing: Algorithms, Architectures, Arrangements, and Applications (SPA). pp. 195–200, 2017. [Google Scholar] [Crossref]
32.
D. A. Muhtasim, M. I. Pavel, and S. Y. Tan, “A patch-based CNN built on the VGG-16 architecture for real-time facial liveness detection,” Sustainability, vol. 14, no. 16, pp. 1–11, 2022. [Google Scholar] [Crossref]
Cite this:
APA Style
IEEE Style
BibTex Style
MLA Style
Chicago Style
Ayanwola, T., Oludele, A., & Agbaje, M. (2023). Enhancing Face Spoofing Attack Detection: Performance Evaluation of a VGG-19 CNN Model. Acadlore Trans. Mach. Learn., 2(2), 84-98. https://doi.org/10.56578/ataiml020204
T. Ayanwola, A. Oludele, and M. Agbaje, "Enhancing Face Spoofing Attack Detection: Performance Evaluation of a VGG-19 CNN Model," Acadlore Trans. Mach. Learn., vol. 2, no. 2, pp. 84-98, 2023. https://doi.org/10.56578/ataiml020204
@research-article{Ayanwola2023EnhancingFS,
title={Enhancing Face Spoofing Attack Detection: Performance Evaluation of a VGG-19 CNN Model},
author={Thomas Ayanwola and Awodele Oludele and Michael Agbaje},
journal={Acadlore Transactions on AI and Machine Learning},
year={2023},
page={84-98},
doi={https://doi.org/10.56578/ataiml020204}
}
Thomas Ayanwola, et al. "Enhancing Face Spoofing Attack Detection: Performance Evaluation of a VGG-19 CNN Model." Acadlore Transactions on AI and Machine Learning, v 2, pp 84-98. doi: https://doi.org/10.56578/ataiml020204
Thomas Ayanwola, Awodele Oludele and Michael Agbaje. "Enhancing Face Spoofing Attack Detection: Performance Evaluation of a VGG-19 CNN Model." Acadlore Transactions on AI and Machine Learning, 2, (2023): 84-98. doi: https://doi.org/10.56578/ataiml020204
cc
©2023 by the author(s). Published by Acadlore Publishing Services Limited, Hong Kong. This article is available for free download and can be reused and cited, provided that the original published version is credited, under the CC BY 4.0 license.