Critical State Detection for Adversarial Attacks in Deep Reinforcement Learning

Conducted research on enhancing the robustness of Deep Reinforcement Learning (DRL) agents against adversarial attacks, recognizing the vulnerability of these models in critical applications. Proposed statistical and model-based approaches to identify key states within an episode, demonstrating that targeting less than 1% of states could decrease agent performance by over 40%. Developed a long-term impact classifier to efficiently identify critical states, achieving an 80.3% reduction in average computation time compared to previous methodologies.

2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA)

10.1109/ICMLA52953.2021.00279

Adversarial Attack Analysis on Autonomous Vehicle Steering Prediction

Conducted research on adversarial attacks in autonomous vehicle systems, focusing on steering angle prediction using a nine-layer convolutional neural network. Simulated diverse driving scenarios in Unity to evaluate the impact of adversarial attacks. Under FGSM, observed steering angle deviations of up to 6°, with a 12% reduction in lane-keeping performance. Patch-based attacks resulted in a 15% increase in collision rates.

Yet to be Published

KnowMeBot

I developed a chatbot using OpenAI and LangChain to simulate my persona and answer questions from a custom dataset. By leveraging Retrieval-Augmented Generation (RAG), the chatbot delivers context-based, high-quality interactions. Deployed using Flask, it provides a seamless real-time chat interface with robust handling of out-of-scope queries, ensuring an engaging and reliable user experience.

Learn More

VisionBridge

I developed a bi-directional content accessibility tool designed to bridge the gap between sighted and blind users by converting text to Braille and Braille to text. The tool supports seamless interaction through both audio and Braille formats, ensuring inclusivity. It incorporates a CNN-based Braille character classification model, achieving 97% accuracy, making the solution highly scalable and enhancing real-time accessibility for diverse user needs.

Learn More

Flight Delay Predictor

Developed a predictive model to forecast flight delays, combining binary classification to identify delayed flights and regression analysis to estimate delay duration. The model utilized a Random Forest Classifier and Gradient Boosting Regressor, achieving high accuracy by addressing data imbalance with SMOTE and integrating comprehensive flight and weather datasets. This solution enhances operational efficiency by providing actionable insights for airlines and passengers.

Learn More