close

Вход

Забыли?

вход по аккаунту

?

978-3-319-69923-3 13

код для вставкиСкачать
Content-Independent Face Presentation Attack
Detection with Directional Local Binary
Pattern
Le Qin1, Le-Bing Zhang1, Fei Peng1(&), and Min Long2
1
College of Computer Science and Electronic Engineering, Hunan University,
Changsha 410082, China
{qinle,zhanglebing}@hnu.edu.cn, eepengf@gmail.com
2
College of Computer and Communication Engineering,
Changsha University of Science and Technology, Changsha 410014, China
caslongm@gmail.com
Abstract. Aiming to counter photo attack and video attack in face recognition
(FR) systems, a content-independent face presentation attack detection scheme
based on directional local binary pattern (DLBP) is proposed. In order to
minimize the influences of the image content, DLBP is proposed to investigate
the noise characteristics of the facial image. By using directional difference
filtering, the discrepancies between the real face and the facial artefact in terms
of the consistency of adjacent pixels are effectively exploited. With the DLBP
feature, the detection is accomplished by using a Softmax classifier. Experiments are done with four public benchmark databases, and the results indicate its
effectiveness both in intra-database and cross-database testing.
Keywords: Face presentation attack detection
component Directional local binary pattern
Content-independent Noise
1 Introduction
As one of the most important identity authentication mechanisms, biometric identification has been widely used in entry access systems, criminal investigations, and in
high-security inspection equipment. Among them, face recognition (FR) has attracted
extensive attention due to its high security, good stability, and ease of use [1]. However, images or videos containing a target’s face can be easily acquired from online
social networks nowadays. If they are misused by malicious attackers, it is possible to
launch presentation attack (also known as spoofing attack) on FR systems [2].
The existing face presentation attack detection (abbreviated as PAD, also known as
spoofing detection or liveness detection [3]) methods can be classified into sensor-based
methods and feature-based methods. Sensor-based methods [4, 5] can improve the
security of FR systems by using additional equipment, but they increase cost and
complexity of applications. Feature-based methods are mainly focused on motion
analysis or texture analysis of original face images. Static texture features [6–10],
dynamic texture features [11, 12], image quality features [13], motion information [14],
© Springer International Publishing AG 2017
J. Zhou et al. (Eds.): CCBR 2017, LNCS 10568, pp. 118–126, 2017.
https://doi.org/10.1007/978-3-319-69923-3_13
Content-Independent Face Presentation Attack Detection with DLBP
119
and hybrid features [15–18] are generally extracted, used as an input of a classifier to
determine the liveness. Nevertheless, these clues are sensitive to the contents of the
facial image, and cannot be well generalized under different detection scenarios. To
exploit the noise signatures of the facial video, Pinto et al. proposed a face PAD method
based on visual rhythms [19]. The residual noise is first extracted by individual frames,
and their Fourier spectrum are computed. After that, the detection features are extracted
from the visual rhythms by calculating local binary pattern (LBP), gray level
co-occurrence matrix, and histogram of oriented gradient. However, the residual noise is
still inevitable contaminated by the image contents.
To counter photo attack and video attack in FR systems and to address the above
problems, directional local binary pattern (DLBP) is proposed to investigate the noise
characteristics of facial images. This investigation is based on the fact that artefacts in a
photo or video are recaptured by a device while real faces are originally captured
images. Recaptured images tend to be more seriously distorted by the reproduction
process, and thus additional noise components are introduced. As known to us, there
exists inherent discrepancies between the recaptured images and the originally captured
images in terms of noise components. On the basis of these properties, it is reasonable
to assume that the real faces and the artefacts can be identified by different noise
features. DLBP takes advantage of the directional difference filter [20, 21], and it can
capture the essential distinction between the real face and the artefact.
The contributions of this paper are:
(1) In order to minimize the influence of image content, DLBP is proposed to extract
the noise characteristics of facial image, the rationale and the motivation are
explained.
(2) Based on the DLBP, a novel face PAD method is proposed, and it can achieve
stable performance across four public benchmark databases.
2 Directional Local Binary Pattern
2.1
Motivation
The motivation of the proposed scheme is that there exists inherent noise discrepancies
between the real faces and the artefacts (photo attacks and video attacks) considering
from the originally captured images and the recaptured images. As the real face images
belong to the originally captured images, they have strong consistency between adjacent pixels. This property means that the intensities of adjacent pixels are very close,
and the difference between adjacent pixels is approximately independent of the pixel
itself [20, 21]. In contrast, the facial artefacts can be viewed as a type of recaptured
images to a certain extent, which are inevitably affected by the print block effects or
video noises in the recapture procedure. This progress directly imposes on the consistency of the adjacent pixels. Therefore, in order to characterize the facial discrepancies between the real faces and the facial artefacts, DLBP is proposed to capture the
noise features.
120
L. Qin et al.
Figure 1 illustrates the examples of the real faces and the facial artefacts in RGB
color space (a), R channel in RGB color space (b), and the difference filtered image
with the corresponding direction {!, ", %, -} in the R channel (c–f), respectively. It
can be found that the facial texture of the artefact is more complex than that of the real
one. This is mainly because the strong consistency of adjacent pixels in the real facial
image. After directional difference filtering, the correlated image contents are removed,
and only the edges of the real face are retained. While for the facial artefact, the
consistency of adjacent pixels is deteriorated by various noises. Therefore, the impacts
of noise interference are amplified by the directional difference filtering, so the facial
texture of artefact is more complex than that of the real one.
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 1. Examples of images after directional difference filter. From top to bottom: real face, print
attack, and video attack. From the left to the right: (a) RGB images. (b) R channel of RGB images.
(c) R channel images after horizontal directional difference filter. (d) R channel images after
vertical directional difference filter. (e) R channel images after main diagonal directional difference
filter. (f) R channel images after anti-diagonal directional difference filter. (Color figure online)
2.2
Construction of DLBP
For an image I, directional difference filters [20, 21] are performed to get the corresponding directional difference matrix M. Four directional quantities are denoted by
superscripts {!, ", %, -}, respectively. The feature of each direction is first calculated independently, and finally they are concatenated to form the DLBP feature. For
example, the directional difference matrix after difference filtering in horizontal
direction (left-to-right) is defined as:
M ! ðx; yÞ ¼ I ðx; yÞ I ðx; y þ 1Þ;
ð1Þ
Content-Independent Face Presentation Attack Detection with DLBP
121
where I(x, y) 2 {0, 1, 2,…, 255}, and x, y represent the row and the column of I,
respectively. For a given pixel (x, y) of M, its DLBP can be obtained by calculating the
LBP [22] with the corresponding pixel in the directional difference matrix M, and it is
defined as:
DLBPU2
P;R ðx; yÞ ¼
8
P
< P1
:
sðMi Mc Þ2i ;
if
U 2;
ð2Þ
i¼0
PðP 1Þ þ 3;
1; x 0
sðxÞ ¼
;
0; x\0
otherwise,
ð3Þ
where P and R represent the pixel number in the neighborhood and the neighborhood
radius, Mc and Mi represent the intensities of the central pixel and the neighborhood
pixels, respectively. U is the transition time of ‘0’ and ‘1’ (such as ‘01’ or ‘10’) in
DLBP coding mode. The construction of DLBP feature is illustrated in Fig. 2.
Direc onal difference →
Direc onal difference
Direc onal difference ↑
Original image
Direc onal difference
R1
R2
R3
...
R5
R4
...
...
...
DLBP
matrices
...
DLBP features
Direc onal difference
Fig. 2. Construction of directional local binary pattern feature.
3 The Proposed Face Presentation Attack Detection Scheme
On the basis of the different noise characteristics between the real face and the facial
artefact, a face PAD scheme based on DLBP is proposed in this paper (see Fig. 3). For
an input frame (image), the face area is detected and normalized to a facial image with a
fixed size to decrease the computational complexity and to avoid the influence of
different size of the input frame. After that, noise characteristics are extracted by
calculating the proposed DLBP, which are used to characterize the degradation of the
artefact. Finally, these features are fed to a Softmax classifier [23], and the output
predicted label determines the liveness. Given the requirements of real-time detection
in FR scenarios, only the first 25 frames of each video (about 1 s) in the test set are
selected for testing, and the predicted label of the given video is determined by the
average decision score of these 25 frames.
122
L. Qin et al.
Real
So max
Classifier
Direc onal LBP
Input image
or frame
Face detec on and
normaliza on
DLBP feature
Feature extrac on
A ack
Classifica on Final decision
Fig. 3. Framework of the proposed face presentation attack detection scheme.
4 Experimental Results
4.1
Experimental Setup
Experiments are made with four public benchmark databases, and they are ReplayAttack database [24], CASIA face anti-spoofing database (CASIA FASD) [25], MSU
mobile face spoofing database (MSU MFSD) [13], and Replay-Mobile database [26].
For fair comparison, all experiments are strictly performed with the original protocol.
The details of the databases and the evaluation measures such as equal error rate (EER),
half total error rate (HTER), attack presentation classification error rate (APCER), bona
fide presentation classification error rate (BPCER), and average classification error rate
(ACER) can be found in [13, 24–26].
In face detection and normalization, eyes locations provided by [13] are used for
MSU MFSD, CASIA FASD, and Replay-Attack databases. As no public eye location
data are provided in Replay-Mobile database, Face++ SDK [27] is used for eyes
localization in this database. To maintain comparability with the previous works, the
face image is resized to 64 64 before feature extraction. For DLBP, the pixel number
in the neighborhood and the neighborhood radius are set as P = 8 and R = 1, and
uniform pattern (U <= 2) is utilized. Inspired by [6], DLBP features are extracted from
HSV and YCbCr color spaces to exploit color information of facial image.
4.2
Performance Comparison with LBP Variants
Experiments are conducted to compare the performances of the existing LBP variants
with the proposed DLBP. For fair comparison, the LBP variants are extended to
multi-channel version with HSV and YCbCr color spaces, and only features with single
scale are used for comparison. The results are listed in Table 1, and DLBP can achieve
stable detection performance across five protocols.
4.3
Performance Comparison
Experiments have been conducted to compare the performances of the existing face
PAD methods with the proposed scheme, and the results of intra-database testing are
listed in Tables 2 and 3, respectively. The proposed method outperforms the existing
methods with MSU MFSD and Replay-Mobile databases, and also achieves competitive results with CASIA FASD and Replay-Attack databases. However, the EER
Content-Independent Face Presentation Attack Detection with DLBP
123
Table 1. Performance comparison of DLBP and existing LBP variants with Replay-Mobile
database. The results are reported for each protocol: MP – mattescreen-photo, MV – mattescreenvideo, PF – print-fixed, PH – print-hand, and GT – Grandtest.
Method
Test BPCER
MP MV
LBP [22]
12.73 12.73
CoALBP [28]
3.64 0.00
PRICoLBP [29]
1.82 0.00
OC-LBP [30]
26.36 14.55
DCP [31]
1.82 9.09
DLBP (proposed) 6.36 2.73
(%) at APCER = 0.01
PF
PH GT
Mean
16.36 2.73 22.73 13.46
19.09 0.00 24.55 9.46
1.82 0.00 27.27 6.18
22.73 0.91 36.36 20.18
23.64 0.91 10.91 9.27
9.09 0.00 9.09 5.45
Table 2. Performance comparison of intra-database testing with Replay-Mobile database using
frame based evaluation. The results are reported for each protocol: MP – mattescreen-photo, MV
– mattescreen-video, PF – print-fixed, PH – print-hand, and GT – Grandtest.
Method
Test HTER (%)
MP MV PF
IQM [26]
7.70 13.64 4.22
Gabor [26]
8.64 9.53 9.40
DLBP (proposed) 2.18 2.88 3.62
PH
5.43
8.99
0.55
Test ACER (%)
GT
7.80 13.64
9.13 9.53
4.74 5.04
Table 3. Performance comparison of intra-database testing with MSU MFSD, CASIA FASD,
and Replay-Attack databases, respectively.
Method
Frame based evaluation
IDA [13]
GIF + IQA [7]
Pulse + LBP-ms-color [15]
LBP + GS-LBP [10]
DLBP (proposed)
Video based evaluation
LBP-TOP [11]
Noise signatures [19]
Spectral cubes [14]
Color LBP [6]
LDP-TOP [12]
LBP + GS-LBP [10]
DLBP (proposed)
MSU MFSD CASIA FASD Replay-Attack
Test EER (%) Test EER (%) Test HTER (%)
8.58
N/A
7.50
7.87
6.13
N/A
18.70
N/A
3.05
10.23
7.41
1.31
N/A
3.31
6.08
N/A
N/A
N/A
N/A
6.54
8.54
3.33
10.00
N/A
14.00
6.20
8.94
2.53
4.44
7.60
14.27
2.75
2.90
1.75
3.13
4.88
increases with CASIA FASD under frame based evaluation, and this is due to the
presence of additional camera noises in low quality samples.
124
L. Qin et al.
The performance comparison of the cross-database testing [32] is shown in
Table 4. The results indicate that the proposed method can achieve stable performance
even in cross-database testing.
Table 4. Performance comparison of cross-database testing with MSU MFSD, CASIA FASD,
and Replay-Attack databases in terms of mean HTER (%).
Trained on
Tested on
Replay-Attack
CASIA
MSU
FASD
MFSD
60.67
N/A
50.00
N/A
LBP-TOP [32]
Spectral cubes
[14]
Color LBP [6]
35.40
LBP + GS-LBP 40.26
[10]
DLBP (proposed) 46.62
CASIA FASD
Replay-Attack MSU
MFSD
49.81
N/A
34.38
N/A
MSU MFSD
Replay-Attack CASIA
FASD
N/A
N/A
N/A
N/A
32.90
36.07
37.90
48.36
21.00
18.57
44.80
45.31
45.70
40.59
31.08
21.63
26.26
48.84
40.20
To analyze the processing time, the evaluation is implemented in a PC with a
configuration of 2.90 GHz CPU and 16 GB RAM without parallel processing, the
average processing time for the test set of MSU MFSD is 1.277 s per video, which
indicate the potential application in real time detection.
5 Conclusions
A content-independent face PAD method based on DLBP is proposed in this paper
(The implementation is available on https://github.com/pp21/DLBP-for-Face-PAD).
The detection exploits the advantage that DLBP can extract the noise characteristics of
facial image, and the influences of image content are minimized. Experiments have
been done with four public benchmark databases, and the results demonstrate the stable
performance of the proposed PAD scheme both in intra-database and cross-database
testing. The future work will focus on the improvements of sampling and quantization
strategies of the proposed DLBP by using feature driven approaches.
Acknowledgments. This work was supported in part by project supported by National Natural
Science Foundation of China (Grant Nos. 61572182, 61370225), and project supported by
Hunan Provincial Natural Science Foundation of China (Grant No. 15JJ2007).
The authors would like to thank Idiap research institute, Institute of Automation, Chinese
Academy of Sciences (CASIA), and Michigan State University for providing the benchmark
databases. The authors would also like to thank Zi-Xing Lin and Xiang Zhang for their kind
proofreading of this manuscript.
Content-Independent Face Presentation Attack Detection with DLBP
125
References
1. Ramachandra, R., Busch, C.: Presentation attack detection methods for face recognition
systems: a comprehensive survey. ACM Comput. Surv. 50(1), 8 (2017)
2. Li, Y., Li, Y., Xu, K., Yan, Q., Deng, R.: Empirical study of face authentication systems
under OSNFD attacks. IEEE Trans. Dependable Secur. Comput. (2016, in Press). doi:10.
1109/TDSC.2016.2550459
3. Information Technology—Biometric Presentation Attack Detection—Part 1: Framework,
ISO/IEC JTC1 SC37 Biometrics, ISO/IEC Standard WD 30107-1 (2015)
4. Lin, Q., Li, W., Ning, X., Dong, X., Chen, P.: Liveness detection using texture and 3D
structure analysis. In: You, Z., Zhou, J., Wang, Y., Sun, Z., Shan, S., Zheng, W., Feng, J.,
Zhao, Q. (eds.) CCBR 2016. LNCS, vol. 9967, pp. 637–645. Springer, Cham (2016). doi:10.
1007/978-3-319-46654-5_70
5. Biggio, B., Fumera, G., Marcialis, G.L., Roli, F.: Statistical meta-analysis of presentation
attacks for secure multibiometric systems. IEEE Trans. Pattern Anal. Mach. Intell. 39(3),
561–575 (2017)
6. Boulkenafet, Z., Komulainen, J., Hadid, A.: Face anti-spoofing based on color texture
analysis. In: IEEE ICIP, pp. 2636–2640 (2015)
7. Peng, F., Qin, L., Long, M.: POSTER: non-intrusive face spoofing detection based on guided
filtering and image quality analysis. In: Deng, R., Weng, J., Ren, K., Yegneswaran, V. (eds.)
SecureComm 2016. LNICSSITE, vol. 198, pp. 774–777. Springer, Cham (2017). doi:10.
1007/978-3-319-59608-2_49
8. Manjani, I., Tariyal, S., Vatsa, M., Singh, R., Majumdar, A.: Detecting silicone mask based
presentation attack via deep dictionary learning. IEEE Trans. Inf. Forensics Secur. 12(7),
1713–1723 (2017)
9. Boulkenafet, Z., Komulainen, J., Hadid, A.: Face anti-spoofing using speeded-up robust
features and fisher vector encoding. IEEE Signal Process. Lett. 24(2), 141–145 (2017)
10. Peng, F., Qin, L., Long, M.: Face presentation attack detection using guided scale texture.
Multimedia Tools Appl. (2017, in Press). doi:10.1007/s11042-017-4780-0
11. de Freitas Pereira, T., Komulainen, J., Anjos, A., De Martino, J.M., Hadid, A., Pietikäinen, M.,
Marcel, S.: Face liveness detection using dynamic texture. EURASIP J. Image Video Process.
2014(1), 1–15 (2014)
12. Phan, Q.T., Dang-Nguyen, D.T., Boato, G., De Natale, F.G.: Face spoofing detection using
LDP-TOP. In: IEEE ICIP, pp. 404–408 (2016)
13. Wen, D., Han, H., Jain, A.K.: Face spoof detection with image distortion analysis. IEEE
Trans. Inf. Forensics Secur. 10(4), 746–761 (2015)
14. Pinto, A., Pedrini, H., Robson Schwartz, W., Rocha, A.: Face spoofing detection through
visual codebooks of spectral temporal cubes. IEEE Trans. Image Process. 24(12), 4726–
4740 (2015)
15. Li, X., Komulainen, J., Zhao, G., Yuen, P.C., Pietikäinen, M.: Generalized face antispoofing by detecting pulse from face videos. In: IEEE ICPR, pp. 4244–4249 (2016)
16. Yang, D., Lai, J., Mei, L.: Deep representations based on sparse auto-encoder networks for
face spoofing detection. In: CCBR, pp. 620–627 (2016)
17. Wu, L., Xu, Y., Xu, X., Qi, W., Jian, M.: A face liveness detection scheme to combining
static and dynamic features. In: You, Z., Zhou, J., Wang, Y., Sun, Z., Shan, S., Zheng, W.,
Feng, J., Zhao, Q. (eds.) CCBR 2016. LNCS, vol. 9967, pp. 628–636. Springer, Cham
(2016). doi:10.1007/978-3-319-46654-5_69
126
L. Qin et al.
18. Patel, K., Han, H., Jain, A.K.: Cross-database face antispoofing with robust feature
representation. In: You, Z., Zhou, J., Wang, Y., Sun, Z., Shan, S., Zheng, W., Feng, J.,
Zhao, Q. (eds.) CCBR 2016. LNCS, vol. 9967, pp. 611–619. Springer, Cham (2016). doi:10.
1007/978-3-319-46654-5_67
19. Pinto, A., Schwartz, W.R., Pedrini, H., de Rezende Rocha, A.: Using visual rhythms for
detecting video-based facial spoof attacks. IEEE Trans. Inf. Forensics Secur. 10(5), 1025–
1038 (2015)
20. Pevny, T., Bas, P., Fridrich, J.: Steganalysis by subtractive pixel adjacency matrix. IEEE
Trans. Inf. Forensics Secur. 5(2), 215–224 (2010)
21. He, Z., Lu, W., Sun, W., Huang, J.: Digital image splicing detection based on Markov
features in DCT and DWT domain. Pattern Recogn. 45(12), 4292–4299 (2012)
22. Ojala, T., Pietikäinen, M., Mäenpää, T.: Multiresolution gray-scale and rotation invariant
texture classification with local binary patterns. IEEE Trans. Pattern Anal. Mach. Intell. 24
(7), 971–987 (2002)
23. Anzai, Y.: Pattern Recognition and Machine Learning. Elsevier, Amsterdam (2012)
24. Chingovska, I., Anjos, A., Marcel, S.: On the effectiveness of local binary patterns in face
anti-spoofing. In: IEEE BIOSIG, pp. 1–7 (2012)
25. Zhang, Z., Yan, J., Liu, S., Lei, Z., Yi, D., Li, S.Z.: A face antispoofing database with diverse
attacks. In: IAPR ICB, pp. 26–31 (2012)
26. Costa-Pazo, A., Bhattacharjee, S., Vazquez-Fernandez, E., Marcel, S.: The REPLAYMOBILE face presentation-attack database. In: IEEE BIOSIG, pp. 1–7 (2016)
27. Face++ SDK. https://www.faceplusplus.com.cn/face-detection/
28. Nosaka, R., Ohkawa, Y., Fukui, K.: Feature extraction based on co-occurrence of adjacent
local binary patterns. In: Ho, Y.-S. (ed.) PSIVT 2011 Part II. LNCS, vol. 7088, pp. 82–91.
Springer, Heidelberg (2011). doi:10.1007/978-3-642-25346-1_8
29. Qi, X., Xiao, R., Li, C.G., Qiao, Y., Guo, J., Tang, X.: Pairwise rotation invariant
co-occurrence local binary pattern. IEEE Trans. Pattern Anal. Mach. Intell. 36(11), 2199–
2213 (2014)
30. Zhu, C., Bichot, C.E., Chen, L.: Image region description using orthogonal combination of
local binary patterns enhanced with color information. Pattern Recogn. 46(7), 1949–1963
(2013)
31. Ding, C., Choi, J., Tao, D., Davis, L.S.: Multi-directional multi-level dual-cross patterns for
robust face recognition. IEEE Trans. Pattern Anal. Mach. Intell. 38(3), 518–531 (2016)
32. de Freitas Pereira, T., Anjos, A., De Martino, J.M., Marcel, S.: Can face anti-spoofing
countermeasures work in a real world scenario? In: IEEE ICB, pp. 1–8 (2013)
Документ
Категория
Без категории
Просмотров
0
Размер файла
793 Кб
Теги
978, 69923, 319
1/--страниц
Пожаловаться на содержимое документа