KAS Lab · Rochester Institute of Technology
As a Ph.D. Graduate Research Assistant in the Khojasteh Autonomous Systems Laboratory (advised by Dr. Mohammad Javad Khojasteh), I study how autonomous robot teams coordinate under constrained energy, communication, and trust for disaster search-and-rescue, alongside inference-time defenses that protect deployed classifiers from unauthorized distillation.
01 / Themes
Core questions guiding current Ph.D. work and collaborative studies.
Primary Ph.D. Focus · KAS Lab
How can autonomous robot teams coordinate reliably when energy, bandwidth, and trust are constrained?
Developing multi-agent reinforcement learning and coordination methods for disaster search-and-rescue teams—spanning dynamic resource allocation, communication relays under bandwidth limits, aerial UAV battery recharging (Fontus), and containment of compromised agents.
ADS-C · Matched-Cost Benchmark
How can deployed classifiers deter unauthorized model extraction without altering legitimate predictions?
Designing closed-form per-input confidence-margin perturbations (ADS-C) and matched-cost evaluation wrappers that preserve 100% of served top-1 predictions while substantially degrading surrogate student distillation.
JAX / Flax NNX · 34-LLM Study
How do we distill compact language models and evaluate alignment for structured decision-making?
Exploring two-stage knowledge distillation for compact local code-generation models in JAX/Flax NNX (codriver, 260M parameters) alongside theory-grounded behavioral alignment studies across 34 LLMs and 198 human participants.
02 / Highlights
Open-source environments and reproducible evaluation harnesses.
Machine Learning Security · Inference-Time Defense
Margin-aware inference-time defense against soft-label model extraction that preserves 100% of served top-1 predictions while reducing distilled-student accuracy by 29.6 pp on CIFAR-10, 17.4 pp on CIFAR-100, and 13.3 pp on Tiny-ImageNet.
Multi-Agent Reinforcement Learning · Search-and-Rescue Autonomy
Custom PettingZoo Parallel multi-agent environment with vectorized NumPy dynamics and proximity-triggered aerial battery charging; trained parameter-shared IPPO/PPO policies in PyTorch and Ray RLlib and diagnosed a reward-hacking local optimum against an analytical baseline.
03 / Papers
Demonstrated that decentralized LLM teams using an external auction-based conflict-resolution protocol (CBAA) achieve lower breach rates than centralized or independently acting teams, and built a reproducible testbed evaluating coordination under communication constraints, decision authority, and multimodal visual deception.
@misc{ullah2026consensus,
title = {Consensus Within, Competition Between: Decentralized LLM Teams for Perimeter Defense},
author = {Khawaja Abaid Ullah and Ananya Acharya and Mohammad Javad Khojasteh},
year = {2026},
eprint = {2610.41044},
archivePrefix = {arXiv},
primaryClass = {cs.MA},
url = {https://arxiv.org/abs/2610.41044}
}Margin-aware inference-time defense against soft-label model extraction that preserves 100% of served top-1 predictions while reducing distilled-student accuracy by 29.6 pp on CIFAR-10, 17.4 pp on CIFAR-100, and 13.3 pp on Tiny-ImageNet.
@misc{ullah2026adsc,
title = {ADS-C: Antidistillation Sampling for Classification},
author = {Khawaja Abaid Ullah and Mohammad Javad Khojasteh},
year = {2026},
eprint = {2607.15467},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2607.15467}
}First-of-its-kind unified benchmark and matched-cost evaluation protocol for test-time antidistillation defenses in classification, comparing existing defenses under a common threat model for suppressing student distillation.
@unpublished{ullah2026antidistillbench,
title = {AntidistillBench: A Unified Benchmark of Test-Time Antidistillation Defenses for Classification},
author = {Khawaja Abaid Ullah and Mohammad Javad Khojasteh},
year = {2026},
note = {Manuscript under review}
}Evaluated social-judgment reliability and human-LLM alignment across 34 large language models and 198 human participants using theory-grounded persona ratings and controlled analyses of gender presentation.
@misc{mahmud2026socialattraction,
title = {How Do Large Language Models Judge Social Attraction? Evidence from Theory-Grounded Persona Ratings Across Multiple LLMs and Humans},
author = {Hasan Mahmud and Khawaja Abaid Ullah and Mohammad Javad Khojasteh and Jamison Heard and Prabu David},
year = {2026},
eprint = {2608.09717},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2608.09717}
}Interested in collaborating on cooperative search-and-rescue autonomy, ML security, or open-source ML systems?