Research
These are some of the research projects that I have worked on that cut accross domains like Artificial Intelligence, Social Network Analysis and Security and Medicine.
Noisy Speech Simulation with GANs
Speech and Language Lab, Nanyang Technological University, Singapore
-
Speech processing models need to be robust to noise in order to function well in the real world.
-
Attempted to simulate noise with a small volume of noisy training data (~3 minutes).
-
Studied the performance of four types of GANs in simulating four types of noise and developed a comprehensive framework to pick the right kind of GAN for the right type of noise. (GitHub) (Paper)
Automated Cervical Cancer Detection
Manipal Institute of Technology
-
Cervical Cancer is one of the most commonly occurring cancers among women and occurs mainly in low-resource areas.
-
It is one of the most successfully treatable cancers if detected early.
-
Developed a lightweight deep learning-based model to detect abnormalities on the cervix from images obtained from a smartphone-like device.
-
The model is deployed on an Android Device with safety and privacy features.
Political Stance Detection on Twitter
Manipal Institute of Technology
-
Aims to detect the political inclination of a user on Twitter (now 'X').
-
Computed similarity signals to prominent politicians to boost performance to 95.8% accuracy (as compared to 88.61% using tweets).
-
Implemented a vote distribution algorithm to achieve 99% accuracy. (Paper)
Dynamic Friend Grouping in Social Networks
Manipal Institute of Technology
-
Aim to group user relationships into three levels: close friends, friends, and acquaintances. Then, the levels are modified based on user misbehavior in the network.
-
Network interactions and relationships were simulated to overcome dataset issues.
-
Static classification was achieved through clustering algorithms, including DBSCAN, Birch, and Spectral Clustering.
-
Devised and implemented an algorithm to demote relationship status based on misbehavior, taking into account factors like trust, indirect trust, interaction-based friendship, and network standing of users.
Protecting Key Connections and Edge Weights in Social Networks
Manipal Institute of Technology
-
Aims to protect the identification of influential nodes in a network and hide the strength of connections while keeping the published graph data useful for analytics.
-
A graph connection obfuscation algorithm and differential privacy algorithm on edge weights were used. (Paper)
Publications
Study of Generative Adversarial Networks for Noisy Speech Simulation from Clean Speech
Presented at the 2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)
Identity – Attribute Inference in Online Social Network(s) Using Bio-Inspired Algorithms and Machine Learning Approaches
Published in Journal of Mobile Multimedia
Patent
Device and Method for Detecting Cervical Abnormality
Published by Indian Patents Office
Application Number: 202241026524