Nazia Shehnaz Joynab

Graduate Researcher | MS in Computer Science @ UT Dallas

🎓 MS in Computer Science @ University of Texas, Dallas (Spring 2026 - Summer 2027 Exp.); B.Sc. in Computer Science and Engineering from MIST

🔬 Graduate Researcher @ Dr. Latifur Khan's Lab, UT Dallas

📜 2 publications: TRIS, accepted at Findings of ACL: EMNLP 2026, and a Q2 journal paper in Informatics in Medicine Unlocked (link).

👩‍💻 Nearly two years of professional experience as a Software Engineer @SCM_DPI Team, Samsung Electronics (Dec 2023 - Oct 2025).

📭 You can reach me out here.

Hey there! I'm Nazia. Pleased to meet you!

I am currently pursuing my Master of Science in Computer Science at the University of Texas, Dallas (Spring 2026 - Summer 2027 Exp.), where I work as a Graduate Researcher in Dr. Latifur Khan's Lab. Before this, I spent nearly two years as a Software Engineer at Samsung R&D Institute Bangladesh, where I mostly worked on large-scale distributed systems. I obtained my BSc.(Engg.) degree in May 2023 from the department of Computer Science and Engineering (CSE) at Military Institute of Science and Technology (MIST).

My research interests lie at the intersection of AI/ML, NLP, and security, with a particular emphasis on information retrieval and reinforcement learning. Most recently, I contributed to TRIS (Tri-Layer Retrieval Integrity Sieve), a middleware defense against knowledge poisoning attacks in Retrieval-Augmented Generation (RAG) systems, accepted to the Findings of the Association for Computational Linguistics: EMNLP 2026.

Besides this, I have explored the efficacy of Federated Learning in healthcare sector as well. my undergraduate thesis focuses on highlighting potential challenges of classifying heterogenous cancerous cells using federated learning, by experimenting in both IID and non-IID settings, under the supervision of Dr. Muhammad Nazrul Islam (MIST). The findings of this research were subsequently published in Informatics in Medicine Unlocked (link).

At Samsung, I was involved in the development and maintenance of an internal Content Delivery Network (CDN), named Artifact Deployment System (ADS). It is a distributed caching system designed to facilitate the rapid and efficient transmission of large binary artifacts across different global research centers of Samsung. The architectural framework of our system integrates proprietary network accelerators, smart caching mechanisms, proxy configurations, and a public cloud infrastructure to optimize artifact delivery for Samsung's employees and strategic partners.

In my free time, I enjoy travelling and listening to music.

If you require any further information, or want to collaborate on a project, send me an email here.

Research Interests

AI/ML, NLP, Security, Reinforcement Learning, Computer Vision, Robotics

Currently not taking on any new work.

Research

* indicates equal contribution

2026

TRIS: A Tri-Layer Retrieval Integrity Sieve Against Knowledge Poisoning

Munir, M. B., Ononto, A. J., Joynab, N. S., Thuraisingham, B., Khan, L.

Findings of the Association for Computational Linguistics: EMNLP 2026

TL;DR: A tri-layer middleware defense for mitigating knowledge poisoning attacks in Retrieval-Augmented Generation (RAG) systems. We implemented and evaluated retrieval integrity mechanisms using semantic clustering, structural trigger–payload filtering, and LLM consistency verification. Adversarial evaluations across NQ, HotpotQA, and MS-MARCO achieved attack success rates as low as 3% on NQ and 4% on MS-MARCO in the evaluated black-box setting.

2024

A federated learning aided system for classifying cervical cancer using PAP‑SMEAR images

Nazia Shehnaz Joynab, Muhammad Nazrul Islam, Ramiza Rumaiza Aliya, A.S.M. Rakibul Hasan, Nafiz Imtiaz Khan, Iqbal H. Sarker

Manuscript published at Informatics in Medicine Unlocked, 2024

TL;DR: We studied the efficacy of federated learning in cervical cancer prediction on three different experimental settings (2~IID, 1~Non-IID). Our proposed CNN-based FL architecture showed a test accuracy of 94.36% and 78.4% on an IID (Independent and Identically Distributed) and a non-IID setting respectively. In the mentioned IID setting, three clients were distributed images of all the five classes while in the non-IID setting, each client had images of one distinct class only. The results indicate that FL showed a significant performance over traditional ML algorithms.

Projects

Ongoing

MuJoCo simulation: the yellow SO-101 arm grips a black puck on its half of a sling puck board while a simulated human hand pulls the band with a white puck on the other half — MuJoCo match: the SO-101 arm (black pucks) against a simulated human (white pucks).

Sim2Real2Sim Robot Sling Puck on a Low-Cost Arm

CS 6341 Robotics, UT Dallas (Fall 2026)

MuJoCo · PPO (Stable-Baselines3) · OpenCV · SO-101 · LeRobot

A low-cost SO-101 arm learns to play sling puck, a fast two-player game, against a person. The project follows a Sim2Real2Sim approach: develop in simulation, run on the real arm, and use the real failures to improve the simulator.

Simulation: a full-physics MuJoCo match against a simulated human, with scripted grasp, draw, release and block skills.
Learning: a shot-success model trained on 4,000 simulated shots (95.4% accuracy), and a PPO policy that decides when to shoot or block (70% win rate, on par with a greedy planner).
Real arm and vision: arm and board calibration (7.7 mm error), trajectory replay through LeRobot, and an OpenCV tracker that finds all 20 pucks in recorded video at 8 ms per frame.
Real-to-sim loop: designed to use real failures (missed grasps, failed shots) to calibrate the simulator and retrain the policy.

Diagram: a policy and a start frame go into a world model that imagines a rollout; a success head scores it, and the scores rank the policies. Poisoned clips with a trigger enter the world model and the success head, so a failing policy with the trigger ranks first — The threat model: poisoned clips with a trigger enter the world model or the success head, so the judge ranks a failing policy first.

Can a Poisoned Judge Be Caught? Backdoors in World-Model Policy Evaluators

CS 6375 Machine Learning, UT Dallas (Fall 2026)

DINO-WM · Meta-World · MuJoCo · BackdoorBench · BackdoorBox · PyTorch

Can a few poisoned training clips make a pretrained world-model judge pass a failing robot policy, while it still ranks honest policies correctly? And can defenses catch it?

The judge: a pretrained DINO-WM world model imagines Meta-World robot-arm rollouts of about 8 policies per task, and a success head scores them. A linear probe reads the object's pose from the imagined state. The ranking is checked against simulator runs.
Attacks and defenses: 3 public attacks (BadNets, WaNet, GoBA) and 3 public defenses, plus our own action-signature trigger (a quick gripper open–close that changes no pixel) and Physics Cross-Check defense.
Analysis: how a trigger spreads over the imagined horizon, where the backdoor is stored, and how much poison it takes to flip a ranking, on two groups of Meta-World tasks.