Mert Kaya
MenuClose

Research

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)

Accepted, to appear in IEEE Xplore.

The study trains an EfficientNet-B4 deepfake detector on FaceForensics++ under nine augmentation and cutout configurations. It compares AUC, F1, Brier score and log loss, and measures Grad-CAM attention over facial regions. The four metrics disagree on the best configuration: accuracy and probability calibration favor different models.

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 = {UBMK 2026, 11th International Conference on Computer Science and Engineering},
  year = {2026},
  note = {Accepted; to appear in IEEE Xplore}
}

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

Mert Kaya, Advisor: Venera Adanova

MSc thesis, TED University,

Frame-level CNN deepfake detectors compared under augmentation and cutout settings, with Grad-CAM analysis of facial regions.

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}
}

Research areas

Decision models

Models that answer a question with set options and give a calibrated probability for each, evaluated in English, Turkish and German.

Small agent models

Models trained on executed trajectories for agents that share tables, notes and a task board.

Explainable deepfake detection

EfficientNet-B4 detectors, augmentation and cutout experiments, and Grad-CAM analysis of facial regions.

Language models for Quenya and Sindarin

A small model that reads and writes both languages, including Tengwar script, grounded in a grammar-verified corpus.