I am a first-year Ph.D. student at the Illinois Institute of Technology, advised by Prof. Binghui Wang. My research focuses on trustworthy machine learning, with an emphasis on privacy and security in large language models.

Previously, I completed my B.Sc. in Electrical & Electronic Engineering at Bangladesh University of Engineering and Technology (BUET), where I worked on medical AI, model compression, and multimodal learning.

Office: Room 019A Stuart Building, 10 W 31st St, Chicago, IL 60616, US

Illinois Tech Logo BUET Logo
Research Interests
LLM Privacy & Security

Investigating privacy vulnerabilities and security threats in large language models, including membership inference, data extraction, and adversarial attacks.

Privacy-Preserving Machine Learning

Developing systems that enable secure and privacy-aware collaboration across institutions. Contributing to PDASP-T, a holistic privacy-preserving collaborative data sharing system for intelligent transportation.

Medical AI & Multimodal LLMs

Building and compressing vision-language models for medical applications, including dermatological diagnosis and tuberculosis detection, with a focus on efficiency and interpretability.

Publications
Gradient Extrapolation
Gradient Extrapolation-Based Policy Optimization
IN Swapnil, A Saha, TA Khan, MA Haque, SN Lim
arXiv preprint, 2026

Introduces gradient extrapolation techniques for improved policy optimization in reinforcement learning.

Compression Strategies
Compression Strategies for Efficient Multimodal LLMs in Medical Contexts
TA Khan, A Saha, IN Swapnil, MA Haque
arXiv preprint, 2025

Evaluates structural pruning and activation-aware quantization for compressing medical multimodal LLMs, achieving 70% memory reduction with improved accuracy.

CLARIFY
CLARIFY: A Specialist-Generalist Framework for Accurate and Lightweight Dermatological Visual Question Answering
A Saha, TA Khan, IN Swapnil, MA Haque
arXiv preprint, 2025

A specialist-generalist framework for dermatological VQA that balances accuracy with lightweight deployment.

GRPO++
GRPO++: Enhancing Dermatological Reasoning under Low Resource Settings
IN Swapnil, A Saha, TA Khan, MA Haque
arXiv preprint, 2025

Enhances dermatological reasoning capabilities under low-resource constraints through improved training strategies.

Skin Cancer Classification
Skin Cancer Classification Using Pre-trained CNNs: A Transfer Learning Approach Addressing Imbalanced Data Challenges
S Sobhan, A Saha, TA Khan, A Zami
NCIM 2025 (2nd International Conference on Next-Generation Computing, IoT and Machine Learning)

A transfer learning approach using pre-trained CNNs for skin cancer classification, addressing imbalanced dataset challenges.

Tuberculosis Detection
A Multi-Stage Deep Learning Approach to Tuberculosis Detection with Explainable Insights
S Sobhan, A Zami, M Ahmed, TM Zihan, TA Khan, A Saha
NCIM 2025

A multi-stage deep learning pipeline for tuberculosis detection with explainable AI insights for clinical interpretability.

Coming Soon
Projects
AgroBot
AgroBot: AI-Powered Precision Weeding for Sustainable Agriculture in Bangladesh
August 2025

A low-cost autonomous precision weeding robot for small and marginal farmers. Vision-guided weeding with YOLOv5, automated navigation via wheel encoder + IMU with Extended Kalman Filter, and targeted herbicide spraying.

mAP50: 0.763 <0.2ms latency 90% less herbicide ~25,140 BDT
Dermatological Assistant
AI-Powered Dermatological Assistant: Bridging Healthcare Gaps Through Multimodal Intelligence
August 2025

A multimodal framework combining image-based diagnosis with visual question answering, powered by DINOv2 and a compressed LLaVA model. Trained via four stages: auxiliary classification, medical reasoning, interaction optimization, and resource-efficient deployment.

82.05% accuracy 9/10 interaction <4.5 GB memory
Contact

Template adapted from Jon Barron
Last updated June 2026