cv

Applied Scientist Intern

Amazon AWS

Working on security of Code Agents

• Manager: Zijian Wang

Applied Scientist Intern

Amazon AWS

Worked on reliable Code LLM

• Managers: Zijian Wang, Varun Kumar

Graduate Research Assistant

University of Maryland - College Park

Working on the safety and reliability of AI/ML

• Supervisor: Dr. Soheil Feizi

Graduate Research Assistant

Maryland Cybersecurity Center

Worked on Robust-ML-based Malware Detection

• Supervisor: Dr. Tudor Dumitras

Lecturer

United International University

Research Assistant

Bangladesh University of Engineering and Technology (BUET)

Ph.D.

University of Maryland - College Park

Computer Science

• Supervisor: Dr. Soheil Feizi

B.Sc.

Bangladesh University of Engineering and Technology (BUET)

Computer Science and Engineering

• Supervisor: Dr. Atif Hasan Rahman

Almost AI, Almost Human: The Challenge of Detecting AI-Polished Writing

ACL

We evaluated twelve state-of-the-art AI-text detectors using our created APT-Eval dataset of 15K human-written texts refined to varying degrees with AI tools. We found that detectors often classify even minimally polished text as AI-generated and struggle to distinguish levels of AI involvement.

Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text

arXiv

We introduced Adversarial Paraphrasing, a training-free attack framework that universally humanizes any AI-generated text to evade detection more effectively. Our approach leverages an off-the-shelf instruction-following LLM to paraphrase AI-generated content under the guidance of an AI text detector, producing adversarial examples that are specifically optimized to bypass detection.

ML-Based Behavioral Malware Detection Is Far From a Solved Problem

IEEE SaTML

We presented a quantitative study of how sandbox traces differ from real-world ones, and how it impacts machine learning models. We identified this distribution shift and proposed a solution for ML models that boosted the TPR from 14% to 20%@1%FPR.

Fast Adversarial Attacks on Language Models In One GPU Minute

ICML

We proposed a novel approach in adversarial attack on LLMs, namely BEAST, that can jailbreak, cause hallucination, and membership inference attacks. Our approach can find jailbreaking prompts within one minute under a resource-constrained setting.

LLM-Check: Investigating Detection of Hallucinations in Large Language Models

NEURIPS

We introduced efficient techniques that analyze internal states, attention maps, and output probabilities to detect hallucinations from a single response, significantly improving detection performance, while being less computationally expensive than previous methods.

DRSM: De-Randomized Smoothing on Malware Classifier Providing Certified Robustness

ICLR

We are the first to propose certified robustness in the domain of static malware detection from executables. We demonstrated both theoretical and empirical robustness of our proposed DRSM framework. Besides, we published a new benign dataset, named PACE.

MAlign: Explainable Static Raw-byte Based Malware Family Classification using Sequence Alignment

Computers & Security Journal

We proposed a novel approach, namely MAlign, incorporating concepts from Bioinformatics into Malware Security. We developed a static raw-byte-based malware family classifier with better accuracy and robustness. MAlign also provides explainability by relocating the code blocks responsible for malicious attacks.

Dean's Fellowship

University of Maryland

Innovation Fund

ICT Division, Bangladesh

Teaching Assistant

University of Maryland - College Park

• DATA 200

Lecturer

United International University

• Discrete Mathematics
• Data Structures & Algorithms
• Operating Systems