Research

I study whether an AI answer is correct, whether its confidence is reliable and which parts of an image affected a result. The papers, code and released models are linked here.

Publications

Augmentation and Cutout in Deepfake Detection: A Comparative Study of Accuracy, Calibration, and Attention

Mert Kaya, Venera Adanova

UBMK 2026, 11th International Conference on Computer Science and Engineering, Istanbul (IEEE)

Published in IEEE Xplore, , 78-83

A paper comparing nine ways to train a deepfake detector: how accurate it is, how well its confidence matches its results, and which parts of the face it relies on.

Citation (BibTeX)
@inproceedings{kaya2026augmentation,
  author    = {Kaya, Mert and Adanova, Venera},
  title     = {Augmentation and Cutout in Deepfake Detection: A Comparative Study of Accuracy, Calibration, and Attention},
  booktitle = {2026 11th International Conference on Computer Science and Engineering (UBMK)},
  pages     = {78--83},
  year      = {2026},
  publisher = {IEEE},
  doi       = {10.1109/UBMK71847.2026.11745298},
  url       = {https://ieeexplore.ieee.org/document/11745298}
}

Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques

Mert KayaVenera Adanova

MSc thesis, TED University,

The thesis compares ways to train models that detect AI-altered faces in video. It uses heat maps to show which facial regions affected each result.

Citation (BibTeX)
@mastersthesis{kaya2025explainable,
  author = {Kaya, Mert},
  title = {Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques},
  school = {TED University},
  year = {2025},
  doi = {10.5281/zenodo.18998566}
}

Language Preservation Using Agentic AI Architectures: A Tool-Grounded Small Language Model on a Constructed-Language Testbed

Mert Kaya

Public preprint, not peer-reviewed. Eldalambë’s code, data, models and benchmark are private.

This paper studies Eldalambë, an assistant for Tolkien’s invented Elvish languages, Quenya and Sindarin. The adopted version, Pro-Preserve, writes the expected sentence for 40 of 100 test requests, versus 21 for the earlier version, Pro2508, and identifies Tolkien’s words in 277 of 300 cases, versus 259 for Pro2508.

Citation (BibTeX)
@techreport{kaya2026eldalambe,
  author = {Kaya, Mert},
  title = {Language Preservation Using Agentic {AI} Architectures: A Tool-Grounded Small Language Model on a Constructed-Language Testbed},
  institution = {Eschatia Labs},
  year = {2026},
  url = {https://eschatialabs.com/research/eldalambe-2026/},
  note = {Preprint, not peer-reviewed}
}

Research areas

Models for decisions

JevAlt models read a situation, choose one of the given answers and give a probability for each answer. The tests check both the answer and whether that probability matches how often the model is right.

AI task assistants

AI assistants need shared context and limits on their actions. Tholos gives them tables, notes and permission rules on your computer.

Deepfake detection

Deepfakes are videos in which AI has altered or replaced a face. I compare ways to train detectors and use heat maps to show which facial regions affected each result.

Tolkien’s Elvish languages

Eldalambë helps people read and write Tolkien’s invented Elvish languages, Quenya and Sindarin, and their script, Tengwar. The paper is public; the code, data, model files and test items are private.