Source: https://eschatialabs.com/zh/research/msc-thesis/

[返回研究页](https://eschatialabs.com/zh/research/)

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

Mert Kaya

这篇硕士论文研究如何检测视频中经人工智能篡改的人脸。论文比较了使用多样化图像和遮挡面部区域训练 EfficientNet-B4 的方法，再用 Grad-CAM 热力图显示每次预测受到哪些面部区域的影响。

[打开 PDF (17.1 MB)](https://eschatialabs.com/papers/msc-thesis.pdf)

论文正文为英语。

TED University 计算机科学硕士论文，2025 年。CC BY 4.0。

**导师**: Venera Adanova

**日期**: 2025年9月1日

**出版物标识符（DOI）**: [10.5281/zenodo.18998566](https://doi.org/10.5281/zenodo.18998566)

## Abstract

The progress in artificial intelligence has given rise to the generation of deepfake videos that have increasingly convincing details. While such technology is of immense value to entertainment and creative industries, it has, on the other hand, raised serious concerns regarding privacy, security, and public trust. With increasing sophistication in deepfake generation methods, interpretation and explainability of the detection methods have been identified as key research priorities, especially for judicial matters that require transparent evidence under the analysis of legal practitioners. This study examines the impact of preprocessing methods, visual augmentation, and targeted masking on the performance and explainability of CNN-based deepfake detection frameworks. For the experimental setup, a frame-based spatial domain deepfake detection approach is employed. Many variations and sets of combinations were systematically analyzed to test their effect on detection. Model interpretability was assessed using Grad-CAM visualization techniques. By observing decision-making patterns, it was possible to identify configurations that focus on the manipulated regions. Therefore, this indicates that certain preprocessing configurations enhance accuracy and explainability, paving the way for generating more credible and trustworthy detection pipelines. These results highlight the importance of carefully considering preprocessing techniques in the development of detection methods that can remain effective against the ongoing evolution of deepfake technologies. This work contributes to the development of more explainable and reliable deepfake detection methods that can be implemented in the real world, especially where explainable AI is necessary to establish credibility and legal admissibility.

## 许可与署名

Mert Kaya, Explainable deepfake detection using frame level CNN models: A comparative study of augmentation and cutout techniques, TED University, 2025. Licensed under CC BY 4.0.

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)

## 引用

```
@mastersthesis{kaya2025deepfake,
  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},
  note = {Advisor: Venera Adanova}
}
```

## 链接

- [xdfdet](https://xdfdet.mertkayacs.com/)
- [代码（GitHub）](https://github.com/mertkayacs/xdfdet)
- [模型和数据（Hugging Face）](https://huggingface.co/mertkayacs/xdfdet)
- [UBMK 2026 论文](https://eschatialabs.com/zh/research/ubmk-2026/)
- [Zenodo](https://zenodo.org/records/18998566)