Experiment Configurations for ARC-Net Deepfake Detection Model

A dataset of 500 real images from Bangladesh was combined with a 140k real and fake faces dataset to train and test the ARC-Net model. The model was evaluated on external datasets like the Deepfake Dataset and Deepfake Database, and three out-of-distribution experiments were conducted. The study was authored by Md Shihab Reza and last updated in April 2026.

Use Cases

• Benchmarking deepfake detection models based on reported accuracy, precision, recall, and F1 scores.
• Studying model transferability and out-of-distribution generalization based on cross-dataset experiments described.
• Analyzing the impact of attention-residual mechanisms in neural networks based on the described ablation study.
• Applying Explainable AI (XAI) techniques like Grad-CAM and LIME to interpret model decisions for facial regions.

Strengths

• Model performance metrics are explicitly reported, including 99% accuracy, 1.0 precision, 0.97 recall, and 0.98 F1 score.
• External validation was performed on multiple datasets, and three out-of-distribution experiments were conducted.
• The study includes an ablation study and statistical significance testing using McNemar's test.

Limitations

• The dataset is very small at 5.5 KB, suggesting it contains configuration or summary data, not the primary image dataset.
• Column-level documentation is absent; field semantics must be inferred after download.
• The description focuses on the model and experiment results; the specific contents of the XLS file are not detailed.

Provenance

Source

Md Shihab Reza via figshare

Collection Method

Likely contains experimental parameters, model configurations, or result summaries from the described study.

Freshness

Last updated 2026-04-06 17:29:19

Geography

Includes images from Bangladesh; model tested on South Asian images and other datasets.