THE STUDY BOOKSHELF // KAS LAB · RIT RESEARCH PROGRAM
As a Ph.D. Graduate Research Assistant in the Khojasteh Autonomous Systems Laboratory (advised by Dr. Mohammad Javad Khojasteh), I research multi-agent coordination, dynamic resource allocation, constrained communication, and compromised-agent containment for disaster search-and-rescue robot teams, alongside inference-time defenses against unauthorized model distillation.
VOLUME 01 // Multi-Agent Autonomy & Search-and-Rescue
Multi-Agent RL · Resource Allocation · Compromised-Agent Containment · ROS 2
At RIT’s KAS Lab (advised by Dr. Mohammad Javad Khojasteh), I research multi-agent coordination and decision-making toward autonomous robot teams for disaster search and rescue. Current work spans dynamic resource allocation, multi-agent communication under bandwidth and energy limits, detection and containment of compromised agents, and PettingZoo/RLlib aerial UAV recharging environments (Fontus).
VOLUME 02 // ML Security & Test-Time Antidistillation
Antidistillation · Model Extraction · Margin-Aware Sampling · PyTorch
Prediction APIs return probability vectors that adversaries exploit to train surrogate models via knowledge distillation. With ADS-C (arXiv:2607.15467), we enforce a closed-form per-input confidence-margin budget that preserves every served top-1 prediction 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. In our companion review, directed perturbations under a unified matched-cost wrapper drive distilled students 34.5 pp (CIFAR-10) and 47.7 pp (CIFAR-100) below the hard-label baseline.
VOLUME 03 // LLM Distillation, Evaluation & Optimization
Knowledge Distillation · JAX / Flax NNX · LLM Evaluation · Combinatorial Optimization
My work on language models spans both compact model distillation and rigorous behavioral evaluation. For codriver, I ported Gemma 2 from Flax Linen to the NNX API and applied two-stage distillation (from Gemma 2 2B on text and CodeGemma 2B on fill-in-the-middle code) into a 260M local model. In collaborative research at RIT (arXiv:2608.09717), we evaluated social-judgment reliability and human-LLM alignment across 34 LLMs and 198 human participants.
FRAMED STUDY WALL PLATES
Detailed problem formulations, per-input confidence margin bounds, and multi-agent reinforcement learning diagnostics.
Machine Learning Security · Inference-Time Defense
A margin-aware inference-time defense against model extraction that preserves every served top-1 prediction 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.
−29.6 pp
Multi-Agent Reinforcement Learning · PettingZoo & Ray RLlib
Built a custom PettingZoo Parallel multi-agent environment with vectorized NumPy dynamics, energy-aware transitions, and proximity-triggered charging; trained parameter-shared IPPO/PPO policies in PyTorch and Ray RLlib and diagnosed a reward-hacking “lazy agent” local optimum.
IPPO / PPO
STUDY DESK MANUSCRIPTS
Evaluated social-judgment reliability and human-LLM alignment across 34 large language models and 198 human participants using theory-grounded persona ratings, including 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}
}
Built a unified matched-cost benchmark and a method-agnostic top-1-preservation wrapper; directed perturbations drove distilled students 34.5 pp (CIFAR-10) and 47.7 pp (CIFAR-100) below the hard-label baseline. Reimplemented prior test-time defenses under a common threat model and matched-cost evaluation protocol to isolate which perturbation mechanisms genuinely suppress student learning.
@unpublished{ullah2026reversing,
title = {Reversing the Incentive to Distill: A Review of Test-Time Antidistillation Defenses for Classification},
author = {Khawaja Abaid Ullah and Mohammad Javad Khojasteh},
year = {2026},
note = {Under review}
}
Proposed a margin-aware inference-time defense that preserves every served top-1 prediction 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. Released with an open-source PyTorch evaluation harness across all three benchmarks and multiple student architectures.
@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}
}
Interested in collaborating on multi-agent autonomy, ML security, or open-source ML infrastructure?