SaisanthiyaD.1,*
SuprajaP.1
-
(Department of Networking and Communications, Faculty of Engineering and Technology,
SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu 603203, India
{saisantd, suprajap}@srmist.edu.in)
Copyright © The Institute of Electronics and Information Engineers(IEIE)
Keywords
Emotion recognition, Facial expressions, Electro encephalogram, Collaborative multimodal emotion recognition, Multi-resolution binarized image feature extraction, Dwarf mongoose optimization algorithm, Improved federated learning generative adversarial network
1. Introduction
Emotion is a process that involves expressions, and it operates in both conscious
and unconscious circumstances in humans [1,2]. People communicate with eachother using expressions. Some emotions include sadness,
joy, rage, and fear. In human-computer interactions, more research has been done on
emotion recognition [3]. The system has emotional connections because the human–computer interaction system
is dynamic and complicated. ER is categorized in the brain using electro encephalogram
signals [4-7].Therefore, electro encephalogram signals are a vital part ofresearch [8-10]. The human–computer interaction and multimedia technologies of today are highly advanced,
making emotion recognition automatic [15]. Emotions are adjusted using emotion recognition at the 3rdparty suggestion. Emotional AI is usedin many places, such as health care, entertainment,
and education [16]. Artificial intelligence research in the robotic field has been raised. Many multinational
companies, like Microsoft, Google, Samsung, have investing trillions of dollars in
emotion recognition systems [17,18]. Never the less, more time and domain knowledge are needed to implement this technique.
Thus, ER identifies the user’s emotions and responds with the help of multimedia content
[19,20].
The problem statement of collaborative multi modal emotion recognition is to develop
a system that accurately recognizes emotions from multiple modalities, such as facial
expressions, EEG signals. This is a challenging problem because each modality has
limitations [11-20]. For example, facial expressions can be subtle and difficult to interpret, and EEG
signals can be noisy, making itdifficult to extract features.The motivation forusing
collaborative multimodal emotion recognition through improved federated learning generative
adversarial network (GAN) for facial expressions and EEG signals is to address the
limitations of each modality by combining the strengths of multiple modalities. GANs
are a type of machine learning algorithm that can be used to generate realistic data.
In this case, the GAN can generate synthetic facial expressions or EEG signals similar
to real data. This can be used to improve the performance of the emotion recognition
system by providing more training data and helping to regularize the training process.
Federated learning is a paradigm that allows multiple devices to train a machine-learning
model collaboratively without sharing their data. The combination of collaborative
multimodal emotion recognition, GANs, and federated learning has the potential to
develop more accurate emotion recognition systems.The model upgrades ER accuracy by
combining the strengths of multiple modalities.
Remaining manuscript is organized as follows. The recent research is revealed in part
2. Part 3 outlines the proposed technique, and part 4 reports the outcome and discussions.
Part 5 concludes this manuscript.
2. Literature Survey
Severalstudieshave usedelectroencephalogram-based ER. Among them, some studies arerevised
here.
Zhang et al. [11] presented a Multiple modal ER and EEG-based Cross-subject with Cross-modal Model
(CSCM). The input data wastaken from Standard Energy Efficiency Data (SEED) and SEED-IV.
The presented model achieved abetter area under the curve value and lower f-measure.
Zhao et al. [12] presented an attention-based hybrid deep learning method for EEG emotion recognition.
The model extracted the critical feature details and providedan efficient categorization.
EEG data differential entropy characteristics are extracted and arranged by electrode
location. The typical encoder encodes EEG input and extracts spatial information before
introducing the band attention method to apply adaptable weights to distinct bands.
The LSTM network extracted the temporal features, and the time attention technique
attained critical temporal data. The input data was taken from two datasets, Dataset
for Emotion Analysis using Physiological Signals (DEAP) and SEED. The presented model
achieved better accuracy and lower average running time.
Wang et al. [13] presented the Multi-modal emotion recognition using EEG and speech signal. The Multimodal
Emotion Database (MED4) consisted of four modalities. MED4 is comprised of simultaneously
collected information from individual EEGs, photo plethysmography, speech, and face
pictures after being affected by video stimulus meant to elicit joyful, sad, angry,
or neutral moods. The presented model achieved a lower accuracy.
Wang et al. [14] presented the Multi-Modal Domain Adaptation Variation Autoencoder for EEG-basis ER.
The multimodal domain adaptive variation autoencoder (MMDA-VAE) approach learns shared
cross-domain latent representations of multimodal data. A Multi-modal Variation Auto
Encoder (MVAE) was used to project information from multiple modes into a single space.
The input data was taken from two data sets, SEED and SEED-IV. The presented model
achieves abetter f-measure and lower area under the curve.
3. Proposed Methodology
Emotion identification depends on EEG signals because they can depict variations in
human brain stages. In addition to EEG signals, a face expression signal is used as
an external physiological characterization signal for ER [21,22]. EEG with facial expression signals categorize emotions and are combined to the internal
neural models and external sub-consciousness actions using an improved federated learning
generative adversarial network to make multi-mode emotion recognition. The stability
of emotional expression ability for facial expression and EEG signals throughout time
[23,24]. The ER accuracy is enhanced after the fusion of EEG signals and facial expressions.
Fig. 1 portrays a flowchart of the MER-IFLGAN method.
Fig. 1. Flowchart of the MER-IFLGAN method.
3.1 Data Acquisition
In this study, a video library (VL) is initially made for video emotion-evoked electro
encephalogram experimentations. Ninety video clips found in VL. These video clips
have been combined into the WMV format after being compiled from various films and
TV shows. These 90 video clips span three emotional spectra: pornographic, neutral,
and violent. The pornographic and violent movie clips come from two kinds of films:
action and drama. VL has 90 videoclips, with 30 clips each in pornographic, neutral,
and violent videos. Six examiners (three men and three women) assessed only one emotion
type before involved in the VL. While choosing videoclips, these 6 examiners could
selectavideo clip from the video library if they felt that a particular emotional
type of clip was appropriate. The video emotion-evoked EEG experimentation involved
13 healthy participants, seven men and six women. The subjects are 24–28 years, and
the corrected vision achieved 1.0.
The 90 video clips were considered as the stimulus to elicit EEG signals with various
video emotions. While watching the video clips that were playing constantly on the
computer, the participants wore a 64-lead Quik-Cap electrode cap to produce electro
encephalogram signals that represented various video emotions. A 10–20 system electrode
coordination is used to arrange the electrodes of electrode cap. The EEG signals generated
by experiments are collected and prepared using the Neuroscan system. E-Prime software
invented through PST Company was used to design video emotion-evoked EEG experimentation.
Initially, the computer screens in the presence of the subjects exhibited the instructions
with caution. The subjects began the experimentation by press the space bar after
carefully considering the experimental design and overall experiment content. There
were 90 videos in the library, 10 for each type of emotion, and 30 clips were selected
randomly for each topic.These 30 video clips playback randomly toavert subjects from
making inertial memories. The cross-shape prompt was shown onthe computer screen before
each video clip was played to gain the subjects' attention [25]. After playing each video clip, a rest period was allowed for subjects to remain
silent. The experiment endedafter 30 chosen video clips were all played. The EEG signals
from the participants were recorded at 1000 Hz,a sampling rate through out the experiment.
The test was repeatedfor every subject until the EEG signals of the final 13 subjects
were obtained. Four emotions (happy, sad, fear, and neutral) were considered to examine
how differently EEG and facial expression signals can identify various emotional stages.
3.2 EEG Signal Feature Extraction Utilizing Dwarf Mongoose Optimization Algorithm
The EEG signal is a higher dimension, a weak physiological signal that is non-stationary
and non-linear. The benefits of the feature-selection process using DMO include the
simplicity of understanding and lack of prerequisite knowledge. As a result, DMO can
pick features from complex EEG signals that are both higher dimensional and objectively
more accurate classifications. First, dominant and non-dominant features are extracted
from input electro encephalogram signal feature vector. DMO replicates the compensating
behavioral response of the dwarf mongoose. The efficacy of each solution wascalculated
after the population was initiated in the alpha group. The alpha female was selected
using Eq. (1),
where$fitness_{i}$represent dominant features of EEG Signal. The solutions updating
mechanism is based on Eq. (2)
Let$P_{eep}$represents dominant femalevocalization that keeps the family on track;
$P$represents a distributed random number. The sleeping mound is assessed by Eq. (3)
The scale of the average sleeping was determined using Eq. (4)
Once the baby sitting exchange criteria were satisfied, the approach progressed to
the scouting stage, where the subsequent food source or resting mound was considered.
If the family forages far away in the scout group section, they will find abetter-sleeping
mound [26]. The scout mongoose is expressed as Eq. (5),
where$rand$represents a random number in the range$\left(0,1\right)$;$Cf$ represents
the collective-volatile movement control parameter; $\vec{X}$ represents the movement
vector, which can be determined based on Eq. (6)
This equation removes the non-dominant features as of feature vector of the input
electro encephalogram signal.The dominant features were then combined again using
the feature vectors of the input electro encephalogram signal to produce new feature
vectors.
3.3 Facial Expression Feature Extraction using Multi-resolution Binarized Image-feature
Extraction
The facial expression is an essential portion of expressing emotions. Human faces
and facial expressions can communicate a wide range of emotions. Utilize MBIFE to
identify these states. This stage involves delivering the chosen EEG signal features
to an MBIFE for feature extraction. The central and linear symmetrical were also used.
The histogram image is expressed as Eq. (7),
where$S$is the window size; $R$is the pixel code-word resolution;$P$ is pixel intensity.
The stacked normalized elements are expressed as Eq. (8),
$\delta _{p}(j)$ is expressed as Eq. (9),
The final image depiction was then built through concatenation of histograms acquired
by the application of every filter of multi-resolution bank is expressed as Eq. (10),
where$\mathrm{H }_{{f_{1}}}$represents calculated and concatenated histograms of the
responses acquired with the applied filter bank. The large produced histograms werecollected
column-wise in a single representative matrix, and all the examined defect pictures
of all classes were handled in the same manner, which is expressed as Eq. (11),
where$\mathrm{M }$ isthe processed defect image count [27]. The remaining data reduction and categorization processes are expressed as Eq. (11) as a starting point. The covariance matrix $H^{\mathrm{T}}$ was calculated and is
expressed as Eq. (12)
Using the class difference principle, the between-class scatter matrix is expressed
as Eq. (13)
Eq. (14) expresses the with in-class scatter matrix,
The ratio among the projections of $B_{S}$and $W_{S}$was calculated using the Fisher
criterion approach expressed as Eq. (15),
Certain features were extracted from the EEG signals lacking transformation using
Eq. (15). It incursless computational cost and is simpleto perform. From this, many effectual
features were extracted utilizing Multi-resolution binarized image feature extraction
(MBIFE). Subsequently, the extracted features werefed to the emotion recognition.
3.4 Multimodal Emotion Recognition based Upon Improved Federated Learning Generative
Adversarial Network
The classification procedure was essential for recognizing emotions from facial expressions
and EEG signals. The best classification algorithm should be chosen to develop a clear
and precise mode to forecast emotions in realtime. This determines the efficacy and
precision of multi modal emotion recognition. The Flexible activation Functions with
Improved Federated Learning Generative Adversarial Network (IFLGAN) is proposed. The
IFLGAN classifier was used to identify emotions, such as sadness, fear, happiness,
and neutrality. IFLGAN can tailor to the individual characteristics of emotions and
deal with the complications with facial expression by building various trees levels
over the enormous training dataset utilizing 2 scans, with three times better performance
than existing methods. Little run-time resources are used by IFLGAN, and no storage
space is needed to save temporary data. The emotions are divided into four categories
at the categorization unit: fear, sadness, happiness, and neutral. The Generator $G$
and Discriminator $D$ is expressed in Eq. (16),
where$X$represents the training data along mini-batch size, $G_{1}(Z)$ represents
generated data with mini-batch size [28]. The $soft\max $operation was performed and expressed as Eq. (17),
MMD score is found through computing the predictions, and it is expressed in Eq. (18):
where$G_{i}$ represents the generator and $D_{i}$ represents the generator. The $soft\max
$operation has been performed and expressed as Eq. (19),
where $\mathrm{MMD}_{\mathrm{i}}$ represents the MMD Score. For each generator, the
optimal discriminator is expressed as Eq. (20),
If $\Pr _{i}=PG_{i}$, then ${D}_{i}^{\ast }\left(x\right)=\frac{1}{2}$. The global
minimum of the virtual training criterion is expressed as Eq. (21),
The formula of the global generator is expressed as Eq. (22),
where$\vartheta _{{G_{i}}}$isthe $i^{th}$generator’s parameters;$\vartheta _{{G_{g}}}$denotes
the parameters of the global generator. The parameters of each generatorare replaced
in Eq. (22) using Eq. (23),
where$G_{{g_{a}}}$and $G_{{g_{MMD}}}$is expressed in Eq. (24),
Hence, the IFLGAN is categorized as sad, fear, happy, and neutral.
4. Result and Discussion
This segment defines the experimental out come of the MER-IFLGAN technique. The proposed
approach was simulated in Math Works Inc, MATLAB{\textregistered} version 9.7.0.1190202
(R2019b). The metrics were examined. The obtained results of MER-IFLGAN were compared
with existing MER-CSCM [11] and MER-LSTM [12] models.
4.1 Performance Metrics
The performance of the proposed method was evaluated.
4.1.1 F Measure
This is computed by Eq. (25),
Let$h$specifies true positive;$i$specifies the true negative;$j$specifies false positive.
4.1.2 Accuracy
This was determined using Eq. (26),
where$k$implies false negative.
4.2 Performance Analysis
Tables 1-3 list the efficiency of the MER-IFLGAN technique. In these tables, the performance
metrics were evaluated. The efficiency was compared to the existing MER-CSCM and MER-LSTM
approaches.
Table 1 presents accuracy evaluation. The MER-IFLGAN achieves 31.21% and 34.06% greater accuracy
for happy; 26.01% and 27.79% greater accuracy for sad; 45.34% and 22.78% greater accuracy
for fear; 46.28% and 34.11% greater accuracy for Neutral compared to the MER-CSCM
and MER-LSTM models, respectively.
Table 2 lists the results of the F-measure analysis. The MER-IFLGAN achieved the following
compared to the existing MER-CSCM and MER-LSTM models, respectively:35.67% and 33.54%
better F-measure for happy; 36.73% and 34.71% better F-measure for sad; 43.14% and
46.27% higher F-measure for fear; 45.26% and 45.87% higher F-measure for Neutral.
Table 3 presents the Average Running Time (ART) analysis. MER-IFLGAN achieved the following
compared to the existing MER-CSCM and MER-LSTM models, respectively: 35.45% and 33.32%
shorter ART for happy; 38.77% and 25.89% shorter ART for sad; 35.67% and 45.14% shorter
ART for fear; 42.15%, and 43.26% shorter ART for Neutral.
Fig. 2 presents RoC analysis. The MER-IFLGAN achieved a 2.92% and 4.15% higher AUC value
than the existing MER-CSCM and MER-LSTM methods, respectively.
Table 1. Accuracy evaluation.
|
Methods
|
Accuracy (%)
|
|
Happy
|
Sad
|
Fear
|
Neutral
|
|
MER-CSCM
|
76.48
|
86.61
|
91.78
|
84.34
|
|
MER-LSTM
|
83.51
|
75.31
|
81.73
|
81.79
|
|
MER-IFLGAN (proposed)
|
98.55
|
98.11
|
98.67
|
97.77
|
Table 2. F1 score estimation.
|
Methods
|
F-measure (%)
|
|
Happy
|
Sad
|
Fear
|
Neutral
|
|
MER-CSCM
|
75.76
|
91.67
|
84.64
|
85.33
|
|
MER-LSTM
|
83.45
|
79.35
|
78.78
|
81.71
|
|
MER-IFLGAN (proposed)
|
98.89
|
98.91
|
99.78
|
98.90
|
Table 3. Average Running Time Analysis.
|
Methods
|
ART(s)
|
|
Happy
|
Sad
|
Fear
|
Neutral
|
|
MER-CSCM
|
5.55
|
6.66
|
7.56
|
7.45
|
|
MER-LSTM
|
8.9
|
6.66
|
7.45
|
9.7
|
|
MER-IFLGAN (proposed)
|
2.36
|
3.96
|
4.17
|
5.61
|
5. Conclusion
The Improved Federated Learning Generative Adversarial Network Espoused Multimodal
Emotion Recognition in EEG and facial expression was implemented. The MER-IFLGAN technique
was simulated in MATLAB. The MER-IFLGAN method achieved11.14% and 8.36% higher F-measure
than the existing MER-CSCM and MER-LSTM models, respectively. Most studies used two
distinct emotion signals as their target objects, but the recognized emotion rate
tended to lessen when people delineated obfuscated emotion signals. Hence, the next
study will concentrate on structuring a more efficient emotion database and consolidating
more emotion details to enhance the development of the ER scheme. In addition, there
is a lack of a global public data base that contains video and associated evoked EEG.
A future public video EEG data base can be developed by examining ways to maximize
the video type, count, and length and amassing EEG signals from numerous subjects.
REFERENCES
E.S. Salama, R.A. El-Khoribi, M.E. Shoman, M.A.W. Shalaby, ``A 3D-convolutional neural
network framework with ensemble learning techniques for multi-modal emotion recognition.''
Egyptian Informatics Journal, vol. 22, no. 2, pp. 167-176. 2021.
Y. Wu, J. Li, ``Multi-modal emotion identification fusing facial expression and EEG.''
Multimedia Tools and Applications, vol. 82, no. 7, pp. 10901-10919. 2023.
L. Fang, S.P. Xing, Z. Ma, Z. Zhang, Y. Long, K.P. Lee, S.J. Wang, ``Emo-MG Framework:
LSTM-based Multi-modal Emotion Detection through Electro encephalography Signals and
Micro Gestures.'' International Journal of Human-Computer Interaction, vol. 1, no.
1, pp. 1-17. 2023.
F. H. Shajin, B. Aruna Devi, N. B. Prakash, G. R. Sreekanth, P. Rajesh, ``Sailfish
optimizer with Levy flight, chaotic and opposition-based multi-level thres holding
for medical image segmentation.'' Soft Computing, pp. 1-26. Apr. 2023.
F. H. Shajin, P. Rajesh, M. R. Raja, ``An efficient VLSI architecture for fast motion
estimation exploiting zero motion pre judgment technique and a new quadrant-based
search algorithm in HEVC.'' Circuits, Systems, and Signal Processing, pp. 1-24. Mar.
2022.
P. Rajesh, F. Shajin, ``A multi-objective hybrid algorithm for planning electrical
distribution system.'' European Journal of Electrical Engineering, vol. 22, no. 4-5,
pp. 224-509. Jun. 2020.
P. Rajesh, R. Kannan, J. Vishnupriyan, B. Rajani, ``Optimally detecting and classifying
the transmission line fault in power system using hybrid technique.'' ISA transactions,
vol. 130, pp. 253-264. Nov. 2022.
F.M. Alamgir, M.S. Alam, ``Hybrid multi-modal emotion recognition framework based
on Inception V3 DenseNet.'' Multimedia Tools and Applications, vol. 1, no. 1, pp.
1-28. 2023.
S. Dutta, B.K.. Mishra, A. Mitra, A. Chakraborty, ``A Multi-modal Approach for Emotion
Recognition Through the Quadrants of Valence-Arousal Plane.'' SN Computer Science,
vol. 4, no. 5, pp. 460. 2023.
S. Liu, P. Gao, Y. Li, W. Fu, W. ``Ding, Multi-modal fusion network with complementarity
and importance for emotion recognition.'' Information Sciences, vol. 619, no. 1, pp.
679-694. 2023.
J.M. Zhang, X. Yan, Z.Y. Li, L.M. Zhao, Y.Z. Liu, H.L. Li, B.L. Lu. ``A Cross-subject
and Cross-modal Model for Multimodal Emotion Recognition''. InNeural Information Processing:
28th International Conference, ICONIP 2021, Sanur, Bali, Indonesia, December 8-12,
Proceedings, Part VI 28 2021 (pp. 203-211). Springer International Publishing. 2021.
Z. Zhao, Z. Gong, M. Niu, J. Ma, H. Wang, Z. Zhang, Y. Li. ``Automatic respiratory
sound classification via multi-branch temporal convolutional network'' InICASSP 2022-2022
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
(pp. 9102-9106). IEEE. 2022.
Q. Wang, M. Wang, Y. Yang, X .Zhang. ``Multi-modal emotion recognition using EEG and
speech signals.'' Computers in Biology and Medicine. vol. 149: 105907. 2022.
Y. Wang, S. Qiu, D. Li, C. Du, B.L. Lu, H. He. ``Multi-modal domain adaptation variational
auto encoder for eeg-based emotion recognition.'' IEEE/CAA Journal of Automatica Sinica.
vol. 9(9): 161226. 2022.
M. Maithri, U. Raghavendra, A. Gudigar, J. Samanth, P.D. Barua, M. Murugappan, Y.
Chakole, U.R. Acharya, ``Automated emotion recognition: Current trends and future
perspectives.'' Computer methods and programs in biomedicine, vol. 215, no. 1, p.
106646. 2022.
C. Guanghui, Z. Xiaoping, ``Multi-modal emotion recognition by fusing correlation
features of speech-visual.'' IEEE Signal Processing Letters, vol. 28, pp. 533-537.
2021.
Y. Hu, F. Wang, ``Multi-Modal Emotion Recognition Combining Face Image and EEG Signal.''
Journal of Circuits, Systems and Computers, vol. 32, no. 07, p. 2350125. 2023.
M. Wang, Z. Huang, Y. Li, L. Dong, H. Pan, ``Maximum weight multi-modal information
fusion algorithm of electro encephalographs and face images for emotion recognition.''
Computers & Electrical Engineering, vol. 94, no. 1, p. 107319. 2021.
Y. Zhang, C. Cheng, Y. Zhang, ``Multimodal emotion recognition using a hierarchical
fusion convolutional neural network.'' IEEE access, vol. 9, no. 1, pp. 7943-7951.
2021.
D. Liu, L. Chen, Z. Wang, G. Diao, ``Speech expression multimodal emotion recognition
based on deep belief network.'' Journal of Grid Computing, vol. 19, no. 2, p. 22.
2021.
H. Zhang, ``Expression-EEG based collaborative multimodal emotion recognition using
deep autoencoder.'' IEEE Access, vol. 8, no. 1 pp. 164130-164143, 2020.
F. Aldosari, L. Abualigah, and K.H. Almotairi, ``A normal distributed dwarf mongoose
optimization algorithm for global optimization and data clustering applications.''
Symmetry, vol. 14, no. 5, pp. 1021. 2022.
D. Saisanthiya, P. Supraja. "Heterogeneous Convolutional Neural Networks for Emotion
Recognition Combined with Multimodal Factorised Bilinear Pooling and Mobile Application
Recommendation", International Journal of Interactive Mobile Technologies (iJIM),
2023.
M. Park, and S. Chai, BTIMFL: A Blockchain-Based Trust Incentive Mechanism in Federated
Learning. In International Conference on Computational Science and Its Applications
(pp. 175-185). Cham: Springer Nature Switzerland, vol. 1, no. 1, pp. 1 June. 2023.
H. Zhang, ``Expression-EEG based collaborative multimodal emotion recognition using
deep auto encoder.'' IEEE Access, vol. 8, no. 1, pp. 164130-164143. 2020.
J.O. Agushaka, A.E. Ezugwu, L. Abualigah, ``Dwarf mongoose optimization algorithm.''
Computer methods in applied mechanics and engineering, vol. 391, no. 1, p. 114570.
2022.
L.K.. Pavithra, T. Sree Sharmila, P. Subbulakshmi, ``Texture image classification
and retrieval using multi-resolution radial gradient binary pattern. ``Applied Artificial
Intelligence, vol. 35, no. 15, pp. 2298-2326. 2021.
W. Li, J. Chen, Z. Wang, Z. Shen, C. Ma, X. Cui, ``Ifl-gan: Improved federated learning
generative adversarial network with maximum mean discrepancy model aggregation.''
IEEE Transactions on Neural Networks and Learning Systems. vol. 1, no. 1, pp. 1-12022.
D. Saisanthiya received B.Tech degree in CSE from Arulmigu Meenakshi Amman College
of Engi-neering, Thiruvannamalai affiliated to Anna University ,Tamil nadu at 2009.
M.Tech Degree in CSE from Sastha Institute of Sience and Technology, Chembarambakkam
affiliated to Anna University , Tamil Nadu at 2011. She is currently working towards
the Ph.D. degree at the School of Computing, Faculty of Engineering and Technology,
SRM Institute of Science and Technology, India. His research interests include deep
learning and Machine learning algorithms.
P. Supraja Currently working as an Associate professor, School of Computing, Faculty
of Engineering and Technology, SRM Institute of Science and Technology, Chennai,
Tamil Nadu, India She was a recipient of AICTE Visvesvaraya Best Teacher Award 2020.
Previously She completed the Indo-US WISTEMM Research fellow ship at University of
Southern California, Los Angeles, USA funded by IUSSTF and DST Govt of India and She
served as a Post-Doctoral Research Associate at Northumbria University, Newcastle,
UK and completed her PhD from Anna University in 2017. She has published more than
50 research papers in reputed national and international level journals/conferences.
She received her university-level Best Research Paper Award in 2022 & 2019 also she
has received funding from AICTE for conducting STTP. Her research interests include
Cognitive Computing, Optimization algorithms, Machine learning, Deep Learning, Wireless
Communication, and IoT. She is a reviewer in IEEE, Inderscience, Elsevier and Springer
Journals. She is also a member of several national and international professional
bodies including IEEE, ACM, ISTE, etc. In addition, she has received the young women
in Engineering award and Distinguished Young Researcher award from various International
Organizations.