YMER || ISSN : 0044-0477 http://ymerdigital.

com

Liveness Detection for Face Recognition System

Mrs. Jhansi Lakshmi1*, B. Monica Sri1, P. Harshitha1, P. Pragna1

1
Department of Computer Science and Engineering, Vignan’s Foundation for
Science Technology and Research, Guntur, AP, India
pjl_cse@vignan.ac.in

Abstract
The pandemic state of affairs turns each one’s lives into online mode even though
training had been undertaken via digital assembly sessions. But for a maximum of the
conferences duplication or spoofing of faces turned into traced even for a few secured
conferences in the event that they were given assembly hyperlinks they used to get
connected. But our gadget helps, authenticated customers handiest can login and the
relaxation could not make it possible. In this paper, face popularity is primarily based on
a total identity technique used to offer authentication for the consumer there's an
opportunity for face spoofing consequently the early level of the procedure is to confirm
whether the face is actual or faux then modify the database include the registered one or
not. To deal with this problem, face cascaded characteristics had been calculated and the
usage of deep mastering technique and an answer turned finalized. We use face biometric
authentication in an Open Authorization (OAuth) framework to regulate consistent access
to internet resources. We enforce an entire face verification gadget that includes the
detection of a living face accompanied by face authentication that makes use of Local
Binary Patterns as functions for face popularity. The whole face recognition procedure
includes offerings like a photograph registration service and a face liveness detection
service.

Keywords: Face recognition, deep learning, pandemic, virtual

1. Introduction
A person's face can reveal a lot about their personality and emotional state. Face
recognition is a fascinating and challenging topic with numerous applications, including
law enforcement identification, banking, and security system access verification, and
personal identification [1].
Authentication using facial recognition is susceptible to various attacks, particularly
spoofing attacks. This attack influences the operator by representing the acquisition sensor
in order to breach the system of biometric authentication. Especially, a face-duplication
attack can be carried out by imitating the user with a 2D image, digital video, or 3D mask
and thereby getting access as a valid user [2]. This helps to give the precautions to avoid
spoofing or duplication attacks. Before validating the identification of the face, the
collected measures are used to detect the “liveness”. Those anti-spoofing methods are
divided into two groups such as static techniques and dynamic techniques [3]. The static
approaches work by examining a single 2D static snapshot. The analysis of the temporal
properties of a sequence of input frames is the basis of dynamic approaches.
Implementation of dynamic techniques is slow and hard. Moreover, some dynamic
solutions demand users to follow instructions in order to confirm their existence, but not
all users are willing to do so. The dynamic approaches become disadvantageous as a result
of this. Traditional credentials are used by the OAuth protocol to verify the resource
owner's identity [4], putting the privacy of users' data at risk. Therefore, creating reliable,

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1327

YMER || ISSN : 0044-0477 http://ymerdigital.com

scalable, and maintainable systems has become an essential core of the function of
security. So many safety procedures have been advanced to protect the customers’
identities. These approaches are routinely used in online user authentication to manage
access to users' data and verify their identities. Knowledge-based approaches, on the other
hand, are prone to attacks such as man-in-the-middle, replay, and stolen-verifier assaults
[5]. Based totally techniques are primarily based totally on something the consumer owns,
along with a clever card or a token that may be reused, stolen, or manipulated. In each
expertise and possession technique, the authentication gadget checks what the consumer
is aware of or possesses in place of actually verifying the identification of the requester.
Biometric authentication, on the other hand, checks the identity of requesters by examining
their physiological and/or behavioral characteristics [6].

2. Literature Survey
We have summarized the innovative work from the current literature that is connected
to our given study in this part. In [3], the authors suggested that DT recognition be
approached in an innovative way. In order to blend motion and appearance, a volume LBP
approach was devised. A simplified LBP-TOP operator was also described recently, which
was based on concatenated LBP histograms generated from three orthogonal planes.
Experiments on two DT datasets, as well as comparisons to state-of-the-art findings,
demonstrated that our technique is effective for DT recognition. For the MIT and DynTex
datasets, classification rates of 100% and 95.7% were obtained using VLBP and 100% and
97.1% using LBP-TOP. The authors in [7] seek to advance the state of the art in the field
of 3D mask assaults by assessing spoofing performances on 2.5D and 3D systems, then
analyzing each mask independently with LOOCV, and lastly experimenting on another 3D
mask spoofing database. Face verification studies on 2D, 2.5D, and 3D baseline systems
reveal that they are subject to facial mask spoofing attacks. They also show how the two
types of masks used in the two databases differ. According to the findings, 3D MAD masks
that rebuild face shapes using 2D pictures are less competent than those in the Morpho
database that get facial shapes using a 3D scanner. In [4], the authors offered a thorough
root cause analysis of security risks across the OAuth protocol’s distinct phases. Replay
attacks, network eavesdropping, forced-login CSRF attacks, and impersonation attacks
were discovered by the attacker model to be frequent network assaults that attackers may
exploit to impersonate users and get access to their protected resources.
The authors in [1] use the database search, identifying “matches” or “non-matches”
based on distance or similarity measurements acquired from the pattern matcher, and lastly
making a “accept/reject” decision-based on system policy. Finally, it makes a
“accept/reject” decision depending on system policy. Under such a determination policy,
any user’s identity claim (positive or negative) whose pattern could not be obtained could
be rejected. For an acquired pattern, the policy could declare a match for any distance less
than a fixed threshold and “accept” a user identity claim on the basis of this single match,
or it could declare a match for any distance less than a user-dependent, time-variable, or
environmentally linked threshold and require a user identity claim on the basis of this
single match. In [5], the authors suggest a method that takes into account the effect of a
flashlight on a user’s hair. The efficacy of liveness detection has improved by utilizing a
low-cost auxiliary device, such as a flashlight. Using a dataset encompassing individuals
with haircut-fringe hair, the suggested method is tested and compared to the existing
method. The proposed features produce satisfactory results for the classifiers. The
method’s biggest flaw is that users’ haircuts are restricted. In [6], the authors for the first
time a countermeasure strategy for detecting mask attacks has been proposed. Because the
mask attack database is 2D+3D, the proposed countermeasure can be used on 2D and 3D

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1328

YMER || ISSN : 0044-0477 http://ymerdigital.com

data. An LBP-based method is presented as a countermeasure. On texture pictures, LBP-

based techniques are commonly used. Although the number of research using LBP on
depth maps is increasing, it is still not as frequent as it is with texture pictures. The results
show that the technique works well on texture photos as well as depth maps.
The authors in [8] utilized visible and LWIR pictures to compare and contrast classical
and state-of-the-art appearance-based face recognition systems. Human face LWIR
imagery is not only a genuine biometric, but it is almost certainly superior to equivalent
visible data. The authors in [9], approaches based on Gabor filtering or wavelets detecting
the frequency contents of face image points or regions at various resolutions and
orientations are generally effective, but computationally challenging. Faces can be
considered as a composite of micro-patterns that are well characterized by the operator,
which motivates the use of the computationally simple LBP operator for face description.
Combining multiple texturing methods, such as employing Gabor filtering with LBP,
could be a viable option. The authors in [10], for high-accuracy person recognition,
constructed and evaluated a system that combines facial recognition and speaker
identification modules. To merge the results of our face and speech modules, they
employed a basic Bayes net. Deriving and adding confidence scores that can forecast each
module’s reliability makes this method very effective. The authors in [2], developed a
method for detecting spoofing based on learning. the micro-texture patterns that
distinguish real from false face photos. The suggested method encodes the results
using multi-scale local binary patterns (LBP), which are then fed to a support vector
machine classifier, which detects whether or not there is a living person in front of
the camera. incorporating micro-texture patterns into a feature histogram.
The newly reported technique requires an estimate of fields or other direction-
oriented features on target pictures, but it can significantly improve spoofing
detection accuracy. This appears to apply to a wide range of assault types, as well as
different media and support. Stationarity constraints can be eased by substituting
smaller backdrop windows for bigger background windows, which limit the region
around the face to the lowest set attainable for a specific condition. It is also possible
to exchange processing time for accuracy by restricting the size of the time windows
in which countermeasures are implemented. The tool used to generate the figures in
the work [11] has been made open-source and publicly available such that our results
may be duplicated.
To recognize the eyeblink behavior, the model dependencies among the
observations and states in an undirected conditional graphical framework,
incorporated a new-defined discriminative measure of eye state in order to expedite
inference and convey the most effective discriminative information. The authors in
[12], showed that the proposed approach can achieve great performance with just one
typical web camera under uncontrolled indoor illumination circumstances, even
while wearing glasses is permitted. The approach outperforms cascaded Adaboost
and HMM in comparison experiments. In nature, the proposed eyeblink detection
approach can be used for a variety of applications, including fatigue monitoring,
psychological experimentation, medical testing, and interactive gaming. The authors
in [13], summed up to begin, it introduces REPLAY-ATTACK, a novel spoofing
attack database that includes three different sorts of conceivable attacks that use three
different media and two different recording situations. The database offers a
framework for training, development, and testing, as well as evidence of a baseline
facial recognition system’s vulnerability to attacks. Second, it offers a simple and
repeatable LBP-based face spoofing countermeasure and investigates its
effectiveness against a variety of attacks. The standard LBPu233shows the best

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1329

YMER || ISSN : 0044-0477 http://ymerdigital.com

performance/complexity tradeoff among the LBP variants studied. The authors in

[14], suggested the system’s key feature is that it evaluated the trajectory of numerous
face components utilizing the optical flow of lines. With an equivalent error rate of
0.5 percent on the test data, the liveness detection is successful in distinguishing live
face sequences from still images in motion. The scheme’s resilience was proven by
further examination into actual implementation. Despite being limited to line velocity
estimates, the proposed OFL is capable of providing reliable measurements for face
liveness assessment. The implementation based on Gauss is both efficient and
effective. In the case of face localization, a rapid technique based on optical flow
patterns was also demonstrated to be possible. Furthermore, model-based Gabor
feature classification’s face component recognition is resistant to common sources of
inaccuracy, such as glasses and facial hair.
For the nonlinear diffusion filter of [15], the authors have introduced totally stable
additive operator splitting (AOS) techniques. These schemes meet all of the criteria
for discrete nonlinear diffusion scale-spaces while being simple to implement in any
dimension. In terms of computational and storage effort, the number of pixels has a
linear relationship. AOS schemes on parallel architectures are implemented. These
investigations show that by utilizing the inherent parallelism of AOS schemes, it is
possible to achieve a speedup of an order of magnitude. Face liveness detection is
now possible thanks to a component-based development approach. On three
databases, they tested recognition performance. Various spoofing types present
significant obstacles to earlier techniques in these databases. The most informative
regions are kept, however, thanks to the introduction of H-Face. In the meantime, the
authors in [16], apply Fisher criterion analysis to guide the pooling method, avoiding
interferences between regions with different discriminant abilities. As a result, the
proposed solution improves database performance across the board. The authors in
[17] proposed a solution for modeling the difference in the illumination properties of
live and fake faces, and use the diffusion speed. They proposed using the TV flow
and AOS technique to compute the diffusion speed effectively and in a way that is
resistant to changing illumination conditions. They attempted to encode the local
pattern of diffusion speed values, known as the local speed pattern, to better capture
the difference between real and false faces (LSP). The LSP-based scheme works in
real-time and can thus be used on a variety of mobile devices.
The authors in [18], developed a unique approach for detecting liveness in facial
recognition software that protects against picture spoofing. Based on the Lambertian
model analysis, they investigated the different characteristics of imaging variation
from a living person or a photograph, leading to the design of a novel approach for
utilizing the data contained in the provided image. They proved that an illumination-
invariant face recognition approach may be used to get the necessary latent samples,
allowing us to create a sparse nonlinear/bi-linear discriminative model to distinguish
the inherent surface properties of an image from those of a genuine human face.
Experiments on a large picture imposter database show that the proposed technique
works well for image spoof identification, with the advantages of real-time
verification, non-intrusion, and no extra hardware needs. The authors in [19],
attempts were investigated utilizing high-resolution printed photographs and photos
obtained from LCD panels. Although there are various types of spoofing, we focused
on recognition algorithms that only use one image. Even with dim photos, the
proposed extension method for detecting spoofing attempts under changing
illumination yields good classification results with low false positive and false
negative rates. When compared to the state-of-the-art counterpart, the proposed

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1330

YMER || ISSN : 0044-0477 http://ymerdigital.com

extension lowered the classification error by more than 50% for high-quality printing
spoofs (NUAA Imposter Database) and more than 65 percent for recovered LCD
images. Anti-spoofing is an urgent must with so many devices adopting facial
recognition biometric authentication. The work in [20], proposes combining the CNN
analysis model with the face liveness detection module for image input. The results
of module testing reveal that the system can effectively avoid various sorts of face
spoofing attacks. We put static and dynamic spoof face attacks to the test, including
masks, photo posters, and digital pictures, as well as video replays. Further research
could look into parallel programming techniques that could help facial recognition
programs run faster.

3. Related Work
Face spoofing assaults might be made with the aid of using showing the face the usage
of a show system like a phone or tablet. Efforts at assault which includes this create low-
first-rate face textures and are without problems detected with the aid of using assessing
HSV’s sense and photo first-rate [14]. Color reproduction of display media, like motion
pictures or photos, will likely be constrained as compared to the actual face. Besides, faces
that are represented may also incorporate nearby color versions. The color gamut relies
upon the show media, and friend’s chroma versions might be defined with the aid of using
analyzing the chroma channel’s color feature. It additionally wishes to be tested wherein
the color version gives the maximum precious micro-texture illustration with the aid of
using extracting LBP (Local Binary Pattern) facts from diverse color spaces [11]. This way
is needed to research spoofing in regions with insignificant lighting.
The software program-primarily based on a totally anti-spoofing assault detection
technique has a low fee and better precision that has grown quick withinside the ultimate
numerous decades. Early utility structures require customers to blink, flow their lips or
appear in line with instructions [12] that might successfully reply to print strikes, however
additionally the person revel in is terrible, and it can’t react to video playback assaults.
The researchers commenced investigating an evaluation machine primarily based totally
on a handmade characteristic in coping with those problems. Even though those techniques
can feature properly withinside the selected dataset, they’re now no longer appropriate for
actual-global applications. With the developing variety of public benchmark datasets, a
look at liveness detection has been held, and the accuracy of detection is refreshed
continuously [16].
Compared with all of the primarily based totally software program structures, the
hardware setup technique makes use of a completely unique sensor for image acquisition,
making the distinction between the actual face and the spoofing assaults extra notably, so
the detection impact is substantially extra stable [17]. Due to the distinction in reflectivity
between an actual face and spoofing assault, multi-spectral infrared [8], and faraway Photo
Plethysmography techniques may be hired in liveness detection that has excessive
precision. However, the gathering states are relatively strict, together with the hardware
setup process, which is surprisingly complicated. Therefore, it’s miles tough to be
extensively utilized. Besides representing data, density-primarily based totally gadgets
also are used for liveness detection, a time-of-flight camera. These strategies can
effectively deal with a 2D assault however now no longer 3D [7] ref. offers a machine of
liveness detection with a mild area camera, which may also come across many exceptional
spoofing assaults. However, the moderate area imaging equipment is pricey, and the
invention outcomes are seriously suffering from the mild. In general, hardware-primarily
based totally detection methods’ weak spot is the fee of the device which is steeply priced

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1331

YMER || ISSN : 0044-0477 http://ymerdigital.com

or tough to reap and calls for a further setup process. Therefore, this technique can’t be
broadly applied.

4. Existing System
The overall performance of face recognition devices improved considerably due to
enhancements observed inside hardware and software program strategies withinside the
pc imaginative and prescient field. Whereas, researchers proposed and examined
numerous techniques to defend face reputation structures in opposition to those intrusions.
According to the present strategies, anti-spoofing for face techniques were gathered to
most important groups are hardware primarily based totally method and software program-
primarily based totally method. First, the hardware primarily based totally method calls
for a further tool to hit upon a selected biometric trait including sweat of finger, increase
in heart rate, face thermogram, or mirror image. This tool, included in a biometric
verification device, calls for a person used to hit upon the sign for a residing thing. Few
additional gadgets, including infrared equipment, attain better precision whilst in
comparison to less difficult gadgets. However, additional gadgets were high-priced and
they are hard to implement. Secondly, the software program-primarily based totally
method extracts from the function of the biometric trends via a general sensor for
differentiating between actual and pretend trends. The function extraction takes place after
the biometric trends, including the feel functions withinside the facial image, are received
via way of means of the sensor.
The software program-primarily based totally strategies deal with each of the received
3-D and 2-D trends as 2-D for gathering facts function. By this, the intensity facts are
applied for distinguishing between 3-D stay face and flat 2-D faux face images.

5. Proposed System
The best biometric traits to apply in a specific authentication need to have 5 qualities:
robustness, distinctiveness, availability, accessibility, and acceptability. Robustness is the
shortage of extrade for a person’s function for a period. Distinctiveness is a version of
records about the populace in order that characters may especially recognize. Availability
shows where any customer can use this feature. Accessibility is to benefit from obtaining
the function for the usage of a digital sensor. Accessibility refers back to the attractiveness
of amassing the function of a person. The functions that offer those five features were used
for biometric authentication and Verification machine. Verification is described because
of some same character’s facts to the saved profile, whereas identity refers to where the
joining person’s records fit any person withinside the saved dataset. Before authentication
(verification or identity), only enrolled people are allowed.
For enrollment, the customers were told to expose their behavioral or physiological
traits to the detector. The function record is obtained and exceeded via one in all used
algorithms will tests whether or not the obtained records are real or fake. And, it guarantees
the fine of the picture. The subsequent task is to sign in only the obtained records with the
aid of using appearing localization and alignment. And the obtained records are moved
right into a layout that could be a series of identities saved withinside data.
The biometric system performs four steps in the authentication phase and they are:
1. Data Acquisition: It’s a sensor, like a fingerprint sensor or a web camera, that
gathers biometric data in three different quality levels: low, normal, and high.
2. Preprocessing: It uses noise filters, smoothing filters, and normalizing
procedures to eliminate different data and provide an appropriate set of data.
3. Feature Extraction: Before classifying the obtained data, it will extract
important information.

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1332

YMER || ISSN : 0044-0477 http://ymerdigital.com

4. Classification: In this method, uses the obtained characteristics as input and

allow them to the output labels

6. Architecture
The architecture of the technique considered is shown in Figure. 1. At first, it will detect
whether the person entering into the meeting is real or face. If the person is fake, then it
will not allow him/her into the meeting. If the person is real, then it will check whether the
person is authenticated person or not. If yes, it will allow him into the meeting. If not, then
it will not allow him into the meeting. This process will continue throughout the session.
If any person spoofs his face or any other person enters the meeting behalf of him then it
will detect and shows that person is fake.

Figure 1. Proposed Architecture

7. Methodology
7.1. Convolution Neural Network
We endorse building an anti-spoofing version with principal modules: the liveness
detector and CNN classifier. The scheme for the way our version works is pretty easy. The
enter will skip thru the liveness detection module, with a view to come across eye blinks
or lip actions. If detected, the entry will stay processed to the CNN classifier module for
whether or not the face is faux or actual.
The lifestyles signal detection module at the face is similarly divided into modules: the
blink detection module and the lip movement detection module. For the lip movement
detection module, we use the lip-motion-internet module. The detector module may be run
in actual time on a video document or a webcam’s output. This module detects lip motion
with the aid of using creating a clear out to decide the higher and decrease lips’ places then
calculate the lips separation distance.
To decide if the eyes are blinking or not, we use a module that we’ve got created in our
preceding research. We use an eye-fixed vicinity clear out to discover whether or not the
eyes are blinking or not. Filters are beneficial for detecting the presence of the attention

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1333

YMER || ISSN : 0044-0477 http://ymerdigital.com

vicinity from a person’s face picture graph enter. Once the attention vicinity is captured,
the following step is to use detection for eye openness. In this step, the attention openness
class is applied. This class generates an opportunity of beginning the attention to the
entered image, that’s then analyzed in step with the value.

7.2. Non-linear Diffusion

The main thing for our method is to cope with spoofing assaults the usage of a nonlinear
diffusion primarily based totally on an AOS scheme with a huge time interval to achieve
the pointy edges and maintain the boundary places of the input image. Moreover, we use
a specialized deep convolution neural network structure that can extract complicated and
high-degree capabilities to differentiate the subtle image.

7.3. Additive Operator Splitting

Use non-linear diffusion to detect facial vibrancy to capture sharp edges and preserve
border positions. This allows you to use the input image to quickly distinguish between a
fake image and a real image. This will remove the edges obtained from the flat image. On
the other hand, the actual face remains clear. We used a low-pass filter, the Gaussian
smoothing kernel, and used the scale space parameter (t) to smooth out image noise,
especially in multiscale representations as given by (1).

𝐼𝐼(𝑥𝑥, 𝑦𝑦, 𝑡𝑡) = 𝐼𝐼𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂 (𝑥𝑥, 𝑦𝑦) ∗ 𝐺𝐺(𝑥𝑥, 𝑦𝑦, 𝑡𝑡) (1)

Koenderink eventually showed that convolving a picture at any scale with a Gaussian
function equals the linear diffusion solution, such as the Heat equation as shown in (2).

𝐼𝐼𝑡𝑡 = 𝐼𝐼𝑥𝑥𝑥𝑥 + 𝐼𝐼𝑦𝑦𝑦𝑦 (2)

A generalized diffusion equation is given in (3).

𝜕𝜕𝜕𝜕 = 𝑑𝑑𝑑𝑑𝑑𝑑(𝑔𝑔∇𝐼𝐼) (3)

The diffusion coefficient g is a constant R connected to the speed of the diffusion

process in this case. Linear diffusion has numerous drawbacks, including It involves the
blurring of essential features as well as edge changes while flattening fine scales to coarse
scales. Perona and Malik pioneered the use of a nonlinear diffusion approach based on
partial differential equations (PDEs). This method was dubbed anisotropic diffusion. This
method eliminates the following blurring and localization difficulties that impact linear
diffusion. The non-linear diffusion filter determines edges and uses an explicit scheme to
preserve the position of the edges during the diffusion process which is given in (4).

𝜕𝜕𝜕𝜕 = 𝜕𝜕𝑥𝑥 (𝑔𝑔|∇𝐼𝐼|𝜕𝜕𝑥𝑥 ) + 𝜕𝜕𝑦𝑦 (𝑔𝑔|∇𝐼𝐼|𝜕𝜕𝑦𝑦 ) (4)

However, this scheme has regularization. Weickert has given a semi-implicit scheme to
mark this issue. This scheme works at any timestep size, using the AOS scheme, which
treats the coordinates of all axes the same. The AOS scheme gives fast diffusion even at
large time step size values (for example, it shows the difference between edges on flat and
round surfaces). As shown in Figure 2, smoothing the surface texture of the printed
counterfeit image fades the edges, but the actual image retains the edges and prevents the
spread from spreading. The non-linear diffusion technique for real and forgery images is
depicted in Figure 3.

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1334

YMER || ISSN : 0044-0477 http://ymerdigital.com

(a) The above image depicts the (b) A normalized face with a (c) Diffused picture using AOS
True face and the below image 64 × 64 pixel dimension. (above - 100 iterations, below
a forgery. - 5 iterations).
Figure 2. Non-linear diffusion depicting using different images

(a) The above image (b) Both images depict (c) The above image (d) Both images depict
depicts the True face diffused surfaces scaled depicts the True face diffused surfaces scaled
and the below image from 0 – 255. and the below image from 0 – 255.
forgery. forgery.
Figure 3. Non-linear diffusion for real and forgery images

By estimating the diffusion speed, we extract the information characteristics from the picture
surface as given in (5).

𝐼𝐼(𝑥𝑥, 𝑦𝑦) = | log(𝐼𝐼0 (𝑥𝑥, 𝑦𝑦) + 1) − log(𝐼𝐼1 (𝑥𝑥, 𝑦𝑦) + 1) | (5)

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1335

YMER || ISSN : 0044-0477 http://ymerdigital.com

Where I0 is the original picture and Il is the diluted image. As demonstrated the true picture
surface in Figure 2 has rather sharp edges (e.g., nose and cheek). The borders of the bogus photos,
on the other hand, are smoother. To retrieve the information features, all prior techniques employed
hand-crafted features such as the LBP. This method has certain drawbacks, such as the inability to
extract complicated characteristics. As a result, deep learning using gradient descent is employed
in this study to extract discriminative and greater characteristics from the diffused picture.

8. Results and Discussion

This work proposes a specialized deep convolution neural network that can extract the
complex features of the input diffused image to differentiate between a fake and real face.
Our CNN has proven to be powerful in extracting not only the edges, but also the textures
of the faces. These images are shown in Figure 4. We have achieved the highest reported
accuracy of 99%. Hence, it will show the person is real or fake.

(a) The true face which is a Real face. (b) The Fake face which is a forgery.
Figure 4. Diffused images of real and fake faces
9. Conclusion
With such a lot of gadgets the use of facial reputation biometric authentication, the want
for face anti-spoof is an absolute must. This paper proposes the use of the CNN evaluation
version for photo enter mixed with the face liveness detection module. Based at the effects
of module checking out indicates first rate effects to save you diverse kinds of face spoof
assaults. We take a look at diverse spoof face assaults examined protected static assaults
including masks, image posters or virtual photos, and dynamic assaults including video
replays. Further studies can discover parallel programming strategies that could accelerate
the time for facial reputation programs.

References
11.1. Journal Article
[1] J. Wayman, A. Jain, D. Maltoni, and D. Maio, An Introduction to Biometric Authentication
Systems, Biometric Systems: Technology, Design and Performance Evaluation, 2005, 1-
20.
[2] J. Maatta, A. Hadid, and M. Pietikainen, Face spoofing detection from single images using
micro-texture analysis, 2011 International Joint Conference on Biometrics (IJCB), 2011.
[3] Z. Guoying, and M. Pietikainen, Dynamic Texture Recognition Using Local Binary
Patterns with an Application to Facial Expressions, IEEE Transactions on Pattern
Analysis and Machine Intelligence, 29 (2007), 915-928.

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1336

YMER || ISSN : 0044-0477 http://ymerdigital.com

[4] I. A. Gomaa, G. I. Salama, and I. F. Imam, Biometric OAuth service based on finger-
knuckles, 2012 Seventh International Conference on Computer Engineering & Systems
(ICCES), 2012, 170-175.
[5] L. Weiwen, Face liveness detection using analysis of Fourier spectra based on hair, 2014
International Conference on Wavelet Analysis and Pattern Recognition (ICWAPR), 2014,
75-80.
[6] N. Kose and J. L. Dugelay, Countermeasure for the protection of face recognition systems
against mask attacks, 2013 10th IEEE International Conference and Workshops on
Automatic Face and Gesture Recognition (FG), 2013, 1-6.
[7] N. Erdogmus and S. Marcel, Spoofing Face Recognition With 3D Masks, IEEE
Transactions on Information Forensics and Security, 9 (2014), 1084-1097.
[8] D. A. Socolinsky and A. Selinger, A comparative analysis of face recognition performance
with visible and thermal infrared imagery, 16th International Conference on Pattern
Recognition, textbf4 (2002), 217-222.
[9] M. Pietik¨ainen and A. Hadid, Texture features in facial image analysis, Advances in
Biometric Person Authentication,Springer, 2005, 1-8.
[10] T. Choudhury, B. Clarkson, T. Jebara, and A. Pentland, Multimodal person recognition
using unconstrained audio and video, International Conference on Audio-and Video-
Based Person Authentication Proceedings, 1999, 176-181.
[11] A. Anjos, M. M. Chakka, and S. Marcel, Motion-based counter-measures to photo attacks
in face recognition, Biometrics, IET, 3 (2014), 147-158.
[12] P. Gang, S. Lin, W. Zhaohui, and L. Shihong, Eyeblink-based Anti-Spoofing in Face
Recognition from a Generic Webcamera, IEEE 11th International Conference on
Computer Vision, 2007, 1-8.
[13] I. Chingovska, A. Anjos, and S. Marcel, On the effectiveness of local binary patterns in
face anti-spoofing, 2012 BIOSIG - Proceedings of the International Conference of the
biometrics Special Interest Group (BIOSIG), 2012, 1-7.
[14] K. Kollreider, H. Fronthaler, and J. Bigun, Non-intrusive liveness detection by face
images, Image and Vision Computing, Elsevier, 27 (2009), 233-244.
[15] J. Weickert, B. M. T. H. Romeny, and M. A. Viergever, Efficient and reliable schemes for
nonlinear diffusion filtering, IEEE Transactions on Image Processing, 7 (1998), 398-410.
[16] Y. Jianwei, L. Zhen, L. Shengcai, and S. Z. Li, Face liveness detection with component
dependent descriptor, 2013 International Conference on Biometrics (ICB), 2013, 1-6.
[17] K. Wonjun, S. Sungjoo, and H. Jae-Joon, Face Liveness Detection From a Single Image
via Diffusion Speed Model, IEEE Transactions on Image Processing, 24 (2015), 2456-
2465.
[18] X. 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, Springer, 2010,
504-517.
[19] B. Peixoto, C. Michelassi, and A. Rocha, Face liveness detection under bad illumination
conditions, 2011 18th IEEE International Conference on Image Processing (ICIP), 2011,
3557-3560.
[20] Lin and Su (2019) developed a face anti-spoofing and liveness detection system using CNN
[23]. The image is resized to 256 * 256, and RGB and HSV color spaces are used.

VOLUME 21 : ISSUE 10 (Oct) - 2022 Page No:1337