Parsa Ghazvinian

About me

I’m looking for Summer 2027 research internships. Download my CV or email me.

I’m a Ph.D. candidate in Computer Science at Georgia State University, advised by Dr. Rajshekhar Sunderraman and Dr. Yi Ding (University of Tennessee, Knoxville), and a graduate research assistant at GSU’s CHAI Center. I work on efficient and trustworthy AI: systems that are not only capable, but also computationally efficient, reliable, privacy-preserving and practical to deploy.

A common thread runs through my work: improving AI under real-world constraints, whether that means limited compute, noisy supervision, privacy requirements or undesirable model behavior. My recent work tests what actually makes large models cheaper to adapt and run: a controlled study of rank allocation in LoRA (under review), and an analysis of how visual-token pruning in vision-language models is evaluated. I also study how reliable LLM systems are, including authority bias in LLM-based search engines and learning from noisy, disagreeing human annotators (two EMNLP 2026 papers). Earlier, I worked on making encrypted inference with homomorphic encryption faster and lighter on memory (IEEE TPS, ACNS and IEEE S&P).

More detail on each project is on the Research page.

Parsa on stage in front of the ACM AI Leadership Summit banner
At the ACM AI Leadership Summit, 2026.

Research areas

Efficiency & optimization

Testing what actually makes fine-tuning and inference cheaper, from rank allocation in LoRA to how visual-token pruning in vision-language models is evaluated.

Parameter-efficient fine-tuning Efficient inference Vision-language models Projects and papers

Robust & trustworthy AI

Authority bias in LLM-based search engines, and learning from noisy, disagreeing human annotators.

Reliability of LLM systems Learning from annotator disagreement Projects and papers

Privacy-preserving ML

Making encrypted inference with homomorphic encryption faster and lighter on memory.

Encrypted inference Projects and papers

News

  • 2026Two papers accepted at EMNLP 2026 (Main and Findings): authority bias in LLM-based search engines and noise-robust learning from annotator disagreement.
  • 2026Reviewed for IJCNN 2026 and, as a secondary reviewer, for ACL Rolling Review; served on the PoPETs 2026 Artifact Evaluation Committee.
  • 2025Received the Student Scholarship Award from the Professional Organization of Iranian Americans (POOIA).
  • 2024MOFHEI appeared at IEEE TPS 2024, with a student travel grant. The code is open source.
  • 2024Paper on memory-efficient encrypted inference with homomorphic encryption published at ACNS 2024.
  • 2023Packing-aware pruning for encrypted inference appeared at IEEE S&P 2023 as an extended abstract, with a student travel grant.

Selected publications

The full list is on the Publications page and on Google Scholar.

Authority Bias in Conversational Search Engines for Academic Paper Recommendation

U. Jinadu, P. Ghazvinian, A. Budathoki, B. M. Ampel, R. Sunderraman, Y. Ding

EMNLP 2026 (Main Conference) Robust & trustworthy AI

Line chart from the paper: mean annotator F1 of four models as label noise rises from 0 to 40 percent

When Disagreement Meets Noise: Noise Robust Annotator Embeddings for Subjective NLP

K. Jacob, P. Ghazvinian, U. Jinadu, N. T. Yang, Y. Ding

Findings of EMNLP 2026 Robust & trustworthy AI

Does Learned Rank Allocation Help Shared and Standard Low-Rank Adapters? A Controlled Study at Equal Final Rank Budget

P. Ghazvinian, M. Dhakal, U. Jinadu, Y. Ding, R. Sunderraman

ACL Rolling Review Under review Efficiency & optimization

MOFHEI: Model Optimization Framework for Fast and Efficient Homomorphically Encrypted Neural Network Inference

P. Ghazvinian, R. Podschwadt, M. GhasemiGol, D. Takabi

IEEE TPS 2024 Privacy-preserving ML

Education

Ph.D., Computer Science

Georgia State University, Atlanta, GA

Advisors: Dr. Rajshekhar Sunderraman and Dr. Yi Ding
Expected graduation: December 2027

M.Sc., Computer Science

Georgia State University, Atlanta, GA

Awarded en route to the Ph.D.

M.Sc., Computer Engineering (Software)

University of Zanjan, Zanjan, Iran

Thesis: FDI Attack Detection Using LSTMs in Smart Grids
Ranked 1st in cohort

B.Sc., Computer Engineering (Software)

University of Zanjan, Zanjan, Iran

Thesis: Secure DBaaS Model Using NoSQL Databases
Ranked 2nd in cohort

Research experience

Graduate Research Assistant

INSPIRE Center and CHAI Center, Georgia State University, Atlanta, GA

  • Led a controlled study of adaptive rank allocation in LoRA (under review, ACL Rolling Review): learned allocation was never detectably better than a fixed rank at the same parameter budget.
  • Leading an analysis of how visual-token pruning in vision-language models is evaluated (paper in preparation): the answer-searched “upper bounds” used to judge pruning methods overstate the remaining headroom by about 30 points.
  • Built MOFHEI, an open-source framework that prepares pretrained models for encrypted inference, cutting latency by 9.6× and memory by 4× at negligible accuracy loss (IEEE TPS 2024).
  • Co-developed a caching and dynamic weight-loading scheme that cut the peak memory of encrypted inference by up to 50×, so it can run on low-cost hardware (ACNS 2024).
  • Implemented the noise-robust contrastive objective behind NRA-Embed, which learns from disagreeing annotators without erasing minority perspectives (Findings of EMNLP 2026).

Graduate Research Assistant and Teaching Assistant

Data Mining Lab, University of Zanjan, Zanjan, Iran

  • Built an LSTM-based model that detects false-data-injection attacks in smart-grid telemetry, reaching 96.7% detection accuracy on the IEEE 118-bus test system.
  • Teaching assistant for Machine Learning (45 students per semester): ran lab sessions and wrote assignments.

Honors and awards

Student Scholarship Award

Professional Organization of Iranian Americans (POOIA)

Student travel grants

IEEE S&P (2023) and IEEE TPS (2024)

Academic service

Secondary reviewer, ACL Rolling Review (ARR)

Co-reviewed 2 NLP submissions with the primary reviewer.

Reviewer, International Joint Conference on Neural Networks (IJCNN)

Reviewed 3 submissions.

Artifact Evaluation Committee, Proceedings on Privacy Enhancing Technologies (PoPETs)

Evaluated research artifacts for functionality and reproducibility.

Reviewer, IEEE Transactions on Information Forensics and Security (T-IFS)

Reviewed 2 submissions in cryptography and systems security.