Talk to Me

Shahan Ahmed / Research Notes

Research

Publications, conference presentations, and experimental ML studies organized by research category.

Research index

3

Publications

1

Presentations

3

Studies

1

Initiative

Adversarial Machine LearningNLP & Text ClassificationHealthcare AnalyticsPublic Health DataEmail Security
Category 01

Peer-reviewed publications

Journal articles and manuscript work organized as a clean reading list.

3 items
2023
Journal Article

Synthesis and characterization study

Inorganic Chemistry Communications · Published

2022
Journal Article

Modeling and simulation of solar cells

Optics & Laser Technology · Published

TBD
Journal Article

Federated learning + synthetic data for vaccination equity research

Manuscript in preparation · In preparation

Category 02

Conference presentations

Research presentations and academic posters from symposium settings.

1 item
2024
Conference Presentation

DHS Vaccination Coverage Analysis

2024 MSU Student Research Symposium

Analysis of Demographic and Health Surveys data to identify vaccination coverage gaps across countries.

Category 03

Experimental studies

Applied ML and NLP studies with methods, tools, findings, and result pages.

3 items
Adversarial ML Study2026

Social Engineering & Adversarial Obfuscation in BEC Attacks

Analyzed how Unicode homoglyphs and zero-width characters break keyword-based phishing detection. Built a character-level classifier that flags obfuscated Business Email Compromise emails that evade conventional defenses.

Methods

  • Character-level modeling
  • Unicode homoglyph injection
  • Zero-width character attacks

Tools

PythonScikit-learnNLP

Key result

95.4% detection accuracy

View results
Robustness Evaluation2026

Robustness of Phishing Detection Under Adversarial Unicode Obfuscation

Evaluated whether a TF-IDF + Logistic Regression classifier trained on clean data remains robust when phishing emails are adversarially obfuscated at test time.

Methods

  • Clean vs. adversarial test splits
  • Word-level TF-IDF
  • Logistic Regression

Tools

PythonTF-IDFHugging Face Datasets

Key result

99.8% accuracy · 100% recall

View results
Model Benchmarking2026

OCR Text Classification: DistilBERT vs Longformer-DeBERTa

Benchmarked transformer architectures on OCR-extracted document classification and compared F1 score, inference latency, and memory footprint for production model selection.

Methods

  • F1 / precision / recall comparison
  • Latency profiling
  • Memory footprint analysis

Tools

DistilBERTLongformer-DeBERTaRecharts

Key result

2 architectures benchmarked

View results
Category 04

Active initiatives

Ongoing public-interest data projects and research infrastructure work.

1 item
Active
Open Data Initiative

OpenDataBD

Open data initiative for Bangladesh, building infrastructure to make public government data accessible and usable for researchers and citizens.

Visit