Natasha Sharma
@cuchd.in
Assistant Professor CSE
Chandigarh University
RESEARCH, TEACHING, or OTHER INTERESTS
Artificial Intelligence, Computer Engineering, Computer Science, Computer Vision and Pattern Recognition
Scopus Publications
Scopus Publications
Yuvjeet Arora, Anoop Kumar, Dhuriya Ankit Subhash, Rishav Raj, Sakshi Bajpai, and Natasha Sharma
IEEE
“Progress in facial emotion recognition (FER) is essential for improving interactions in affective computing, social robots, and human-computer interface systems. Although the foundation for FER was established by conventional Convolutional Neural Networks (CNNs), recent developments in computer vision have led to new opportunities using Transformer-based models, which have demonstrated exceptional performance in tasks such as picture classification and semantic analysis. In the framework of an extensive FER dataset, we test the performance of many Vision Transformer (ViT) models in addition to well-known CNNs like MobileNet and DenseNet. With the help of the FER2013 dataset, which is well-known for its varied portrayal of face expressions in grayscale pictures, we used several data augmentation methods to replicate real-world diversity. By using careful training with optimized parameters for both CNN and ViT models, our work provides a comparison analysis that shows the advantages and differences between each model type for FER tasks. This investigation shows that CNNs are still useful for effectively processing face characteristics and supports the viability of ViTs in capturing complex emotional expressions. Our results provide important new understandings of the features of the model that provide precise emotion identification, paving the way for further advancements in emotion-aware systems in a variety of fields.”
Vijay Bhardwaj, Shivam Pandey, Ekta Bhardwaj, and Natasha Sharma
IEEE
Using the examination of Imaging information sets, this paper presents an innovative neural network approach tailored for detecting Alzheimer’s Illness (AI). There are several difficulties in using neuroimaging images examinations to diagnose AI. Because the scans of an AI patient with a normalizing brain show resemblance it can be difficult to diagnose AI in minor to extremely mild stages visually. Prevention for AI is vital. Exact identification of AI is necessary. Recent investigation has demonstrated that machine learning methods are extremely successful than individuals at differentiating between different phases of AI, allowing for timely identification and treatment. The objective of this investigation is to create a novel method that makes use of previously trained deep neural networks in order to reliably identify the various degrees of AI extent, especially when the number as well as calibre of readily accessible information are constrained. This method processes the AI information using an advanced image analysis component before the modelling stage. Applying 4 AI datasets three for the minor, extremely mild, and moderate phases of the illness, in addition to one for the usual stage the suggested approach was contrasted against two popular neural network Visual Geometry Group 16 along with Residual Network 50. This made it possible for us to assess how well the categorization findings worked. Six evaluation standards were employed for evaluating the various approaches. Our technique outperforms any other algorithms currently in use, including a general identification precision of 98.34%, according to the findings.
Dhuriya Ankit Subhash, Natasha Sharma, Anoop Kumar, Sudhanshu Kumar Jha, Ishica, and Rajneesh Pandey
IEEE
The rapid increase in the sharing of fake news poses a major threat in this modern digital world to information integrity. This issue can be addressed by using the Machine Learning (ML) models. This research is based on comparative study of four such ML models- Random Forest Classifier, Logistic Regression, Decision Tree, and Gradient Boosting Classifier. Various performance metrics such as precision, accuracy, recall, Fl-score are employed for model evaluation. The dataset is being used to train the features based on textual content and metadata. This particular analysis produces the results like Random Forest Classifier, which exhibits highest accuracy and provides reliable outcomes in differentiating fake news and actual news. Additionally, the research underscores the importance of data preprocessing, feature selection, and model hyper-parameter tuning in optimising the efficiency in fake news detection. This study provides valuable insights to combat the spreading of fake news in this digital age.
Shivam Pandey, Ekta Bhardwaj, Shikha Gupta, and Natsaha Sharma
IEEE
A central nervous system tumour can result in asymmetrical if it is situated in a region that is affected by the nerves that maintain facial balance. The imbalance is not a symptom of all malignancies in the brain, though. Many tumours might exist in regions which don't influence the proportions of the face or the contour of the forehead. Computerized neural networks are used to analyses scans of brain tissue as well as categories tumours into two categories: benign (not cancerous) or malignant (cancerous) by employing deep learning algorithms. Convolutional Neural Networks (have been employed in the discipline of health care photography for tasks including the categorization of neurological tumours. It is frequently possible to find any present brain tumours by reviewing the individual's magnetic resonance imaging (MRI) results. While confronted with numerous types of neurological tumours as well as an enormous volume of information, this method is laborious and susceptible to mistaken information. The three most common types of brain tumours, glioma, meningioma, and pituitary, were programmed to be detected by deep learning algorithms for the suggested investigation. These algorithms have been optimized employing the Aquila Optimizer (AQO), which at first was employed to earn the preliminary sample subsequent generations along with alterations for the chosen information set, breaking down the results into 70% for the training set and 20% for the set to be evaluated. For the modelling and testing of the brain tumour information set, we employed the AQO optimizer with the VGG-16, VGG19, and Inception-V3 topologies and DenseNet to get the best rate of prediction 97.94 % for the DenseNet algorithm
Randhir Bhandari*, , Digvijay Puri, Natasha Sharma, , and
Blue Eyes Intelligence Engineering and Sciences Engineering and Sciences Publication - BEIESP
In today’s world the computer network communication increases the efficiency most of the organizations. Hence threats have been increased due to these online transactions/communications. These threats necessitate the researchers to improve the existing security protocols and /or develop the new ones. Authentication Protocols are one of the same which can provide the authentication, confidentiality & integrity. For checking the authenticity of messages exchange process in authentication protocols BAN logic is used. The Kerberos encrypt the information for authentications. Many organizations use it and it has five versions and versions 4 and 5 are latest. In one of over previous paper we have generalized the ticket exchange process of version 5. In this paper to make it more authenticated some modifications are proposed to both BAN and Kerberos and we defined them as R- Kerberos & R- BAN. To achieve this, we have added participant’s physical address (MAC Address) as it is unique to every network adapter and can be used as our secret key.