projects

Proposed a unified defense strategy to simultaneously address the security and privacy challenges faced by machine learning models. Designed a novel framework to integrate adversarial training with adversarial watermarks to fortify against evasion attacks, providing SOTA results over the existing techniques. Addressed the privacy concerns raised for the DP‑Adv technique and provided evidence demonstrating its efficacy in maintaining data privacy.

Explored open issues in FL and proposed a state‑of‑the‑art FL algorithm to handle the problem of federated graph clustering. Used power iteration method that employs a subtle strategy to synchronize the eigenvectors among clients. Tested the proposed power iteration approach on the ego‑facebook and email‑Eu‑core dataset and obtained a similarity of 99.8% and 98.85%, comparable to that of global clustering.

Developed two virtual reality (VR) environments ‑ A drawing challenge concentrating on children with autism and a football‑themed balance training challenge that focuses on stroke sufferers. Integrated both the environments with the tracker and calibrated the designed environment with the real‑world task.

Analyzed the pipeline involved in generating an image using scene graphs from textual descriptions of the image. Re‑implemented the process of converting text to a scenegraph and an image to a scenegraph. And then, the generated scenegraphs are passed to a model, which creates an image from the scenegraph. Tested on MSCOCO and Visual Genome dataset using different similarity metrics.

Proposed a novel approach to reduce an image dataset using a Geometrical Homogenous Clustering(GHCIDR). Acquired an accuracy of 99.35 % and 81.10%, and a training data reduction of 87.27% and 32.34%, on MNIST, and CIFAR10 respectively.

Analyzed data collected by extracting tweets containing important trending hashtags about the COVID‑19 pandemic using Twitter API ‑ Tweepy. Used various analytical tools such as clustering, text summarization, semantic analysis, and geobased analysis to get specific and useful insights from the tweets.

Created a command‑line tool for the Linux file system that supports snapshot management Used traditional split‑mirror‑based approach with some improvisation over the data storage required while creating the copy. Implemented various functionalities to create, delete and manage the snapshots of the directories.

Implemented a Mini‑WhatsApp tool that supports features such as chatting(individual/group), file sharing, and profile creation. Understood and implemented Peer‑to‑peer communication for the individual chat and concurrent server for group‑chat. Provided an option to see the status of other online users and request for the group chats.