Experience

Aug 2023 - Present

Computational Data Science Lab for the Web and Social Media, GeorgiaTech, Atlanta, US

• Designing defense techniques for improving the adversarial robustness of BERT, GPT and Llama2 LLMs.
• Creating a universal classifier that leverages LLM embeddings to effectively identify and prevent the generation of harmful content.

May 2023 - Aug 2023

Keysight Labs: Advanced Software Development Center, Atlanta, US

• Enhanced the ML Testing toolbox by developing mutual information based feature visualizations and integrating SHAP for multi-class classification models
• Evaluated the end-to-end development of machine learning models (CNN for 5G Beam Selection, Autoencoder Based Channel Estimator and Equalizer, Error Correction Transformers) using Keysight’s AI-testing pipeline
• Contributed to the AI Testing White Papers on Supervised and Unsupervised learning models

Jan 2023 - May 2023

Institute for People and Technology (IPaT), GeorgiaTech, Atlanta, US

• Developing an Emergency Management Portal using Flask and REACT for real-time tracking and monitoring of flooding in Georgia coastal counties.
• Designing pipelines for processing and storing data in GCP collected from sensors across the Georgia coast.

Sep 2021 - Apr 2022

iTrust Labs, Singapore University of Technology and Design, Singapore

Anomaly Detection in Multi Variate Time Series

• Developed a dual attention (spatial and temporal) based LSTM/GRU model to pre-emptively detect anomalies in a power plant.
• Minimized costs by reducing the false alarm rate to 0.21% with a high detection accuracy of 97.8%.

Semi-Supervised GAN based Trojan Detection

• Researched and implemented a novel semi-supervised GAN for detecting trojaned DNNs.
• Enhanced detection capabilities by integrating a Denoising Autoencoder for attack agnostic one-class training.
• Achieved state-of-the-art performance +3% AUC on computer vision tasks while reducing run-time by 15%.

Sep 2021 - Apr 2022

Fraunhofer IOSB and Karlsruhe Institute of Technology, Germany

• Implemented black-box ZOO and FGSM attacks on DL based Intrusion Detection System. Handled system specific constraints: binary features, one-hot encoded variables, preserving attack capabilities.
• Studied various defense mechanisms including adversarial training, defensive distillation, and denoising autoencoder.
• Integrated Explainable AI tools such as PDP Plots and SHAP to perform understandable and trustworthy decision making.

Jul 2020 - Sep 2020

PRETzel Group, University of Auckland, New Zealand

• Implemented agent(Stochastic Te Punaha Matatini) and differential equation(Modified SEIR, CovidSim 2.0) based covid-19 models using hybrid automata using composition modelling.
• Performed multi-model simulations with different plants and controllers to find the optimal control strategy.
• Analysed the trade off between the economic impact and the health impact by generating a pareto front using ϵ constraint method and particle swarm optimization.