Umid Suleymanov
PhD Student in Computer Science, Virginia Tech
I am a second-year PhD student in Computer Science at Virginia Tech. My research focuses on Trustworthy Machine Learning, specifically investigating LLM Security (jailbreak attacks/defenses), Privacy (membership inference, unlearning), and Few-Shot Learning.
Previously, I was a Data Science Intern at Amazon (AWS) working on proactive security. I was also a two-time finalist at the International Data Analysis Olympiads, obtaining 16th place among 2187 teams.
News
-
Nov 2025
[Paper] Our paper "SIGMA: Search-Augmented On-Demand Knowledge Integration for Agentic Mathematical Reasoning" was accepted at AAAI 2026.
-
Aug 2025
[Paper] Our paper "SSPNet: Semi-Supervised Prototypical Networks for Few-Shot Network Intrusion Detection" was accepted at IEEE MILCOM 2025.
-
May 2025
Joined Amazon (AWS Proactive Security) as a Data Science Intern.
-
Apr 2025
Presented at annual CCI Student Researcher Showcase
-
Dec 2024
Presented at Commonwealth Cyber Initiative (CCI) Southwest Virginia Graduate Student Summit (SWVA)
-
Sep 2024
[Paper] Published research on "Deep Unlearning of Breast Cancer Histopathological Images" and "Contextualized Word Embeddings for Azerbaijani" at IEEE AICT 2024.
-
Aug 2024
Started my PhD in Computer Science at Virginia Tech.
-
Jun 2022
Promoted to Leading Data Scientist at E-Gov Development Center.
-
Nov 2019
Joined E-Gov Development Center as a Senior Data Scientist.
-
Oct 2018
Won 1st Place at the AzInTelecom Hackathon with a solution involving image processing and OCR.
-
Jul 2018
Started as a Data Engineer at E-Gov Development Center.
Selected Publications
Experience
Amazon (AWS Proactive Security)
May 2025 – Aug 2025- Developed statistical and ML models to process large-scale security data.
- Built ETL pipelines, dashboards, and metrics using Python, SQL, and AWS tools.
Virginia Tech
Aug 2024 – Present- Published research on semi-supervised few-shot learning for Network Intrusion Detection.
- Developing policy-grounded RAG-based agentic defenses for LLM jailbreak attacks.
- Researching LLM privacy leakage, membership inference, and memorization analysis.
E-Gov Development Center
Jun 2022 – Jan 2023- Led the design and deployment of end-to-end ML pipelines, integrating predictive models into production systems.
- Developed pretrained language models and word embeddings for Azerbaijani and leveraged them for multiple downstream NLP tasks.
- Built custom Elasticsearch text analyzers that improved search relevance and increased click-through rates by 6%.
Teaching
I have served as an instructor for the following courses:
- Artificial Intelligence University Instructor
- Deep Learning University Instructor
- Intro to Big Data and Analytics University Instructor
- Data and Information Engineering University Instructor
- Instructor @ 30 Days of ML Organized by Google Developer Groups Baku
Academic Service
- ACM Transactions on Privacy and Security Reviewer
- AICT 2025 Conference Reviewer
- AICT 2024 Conference Reviewer
- Intel International Science and Engineering Fair (Local Selection) Judge (CS Domain)
- Intl. Conf. on AI for Digital Governance (AI4DIGIGOV) Tutorial: "Applications of ML in Public Services" (Apr 2021)
Projects
SIGMA: Agentic Math Reasoning
Designed a multi-agent framework that orchestrates specialized agents to perform targeted searches and synthesize findings. Outperforms open and closed-source models on MATH500 and GPQA benchmarks by 7.4%.
SSPNet: Few-Shot NIDS
Proposed a semi-supervised prototypical network for Network Intrusion Detection (NIDS). The model leverages pseudo-labels from high-confidence samples to improve detection of novel attacks with limited labeled data.
Policy-Grounded RAG Defense
Developing policy-grounded RAG-based agentic defenses for detecting and mitigating LLM jailbreak attacks. The system retrieves safety policies dynamically to ground model responses against adversarial inputs.
Agentic Defenses for LLM Privacy
Developing an agentic feedback loop to identify private information leakage in Large Language Model outputs and creating privacy-preserving defenses to mitigate these risks.
Historical Document Segmentation
Developed instance segmentation pipelines (U-Net and Mask R-CNN) to digitize handwritten historical archives. U-Net with Watershed transformation outperformed two-stage architectures.
Deep Unlearning for Histopathology
Applied influence-based unlearning techniques to DenseNet models for breast cancer classification (BreaKHis dataset), enabling data removal compliance while maintaining model utility.
Auditing Fairness-Privacy Trade-offs
Developed a unified empirical framework to audit how fairness-enhancing algorithms influence membership inference privacy risks at the subpopulation level, revealing disparate trade-offs between privacy, utility, and fairness.
Machine Unlearning Quality Score
A systematic metric that quantifies how closely an unlearning method mirrors a fully retrained model without requiring actual retraining.
Contextualized Embeddings (Az)
Trained RoBERTa and GPT-2 models on a 164M token Azerbaijani corpus. This work provided the first high-quality contextual embeddings for the language, significantly improving downstream NLP tasks.
Online News Classification (AZ)
Developed an automated classification system for Azerbaijani news articles using Naive Bayes, SVM, and Neural Networks. Implemented stemming, stop-word removal, and feature reduction to enhance accuracy and performance.
Sentiment Polarity Detection (AZ News)
Built sentiment analysis models for Azerbaijani social news articles using SVM, Random Forest, and Naive Bayes with TF-IDF and frequency-based BOW. Evaluated on a dataset of 30k manually labeled articles; SVM achieved top performance.
End-to-End MLOps Pipeline
Implemented a full ML lifecycle project using MLflow for experiment tracking and model registry, and Docker for containerization, following industry standard data science processes.
Song Recommendation System
Built a scalable recommendation engine using Apache Spark and MLlib. The system utilizes implicit feedback from user history logs to predict and recommend new songs.
Wikipedia Parser & Analyzer
Developed a heuristic-based parser to clean and analyze Azerbaijani Wikipedia data. Analyzed property distributions in templates and external reference usage.