ARC-Net Deepfake Detection Model Configurations and Experimental Results

Bangladesh contributes 500 real facial images to a combined dataset of over 140,000 real and fake faces used for model training. The dataset supports the ARC-Net model, achieving 99% accuracy in deepfake detection, as published by author Md Shihab Reza in April 2026. It documents the experimental setup, performance metrics, and out-of-distribution tests for the proposed detection architecture.

Use Cases

• Benchmark model performance using reported accuracy, precision, recall, and F1 scores against four pre-trained deep learning models.
• Analyze the impact of incorporating Bangladeshi facial images on false positive rate reduction and probability calibration in out-of-distribution experiments.
• Replicate the ablation study by examining the systematic removal or modification of attention and residual convolutional components within ARC-Net.
• Apply McNemar's statistical test to assess significance between competing classifiers using the provided experimental results.
• Utilize Grad-CAM and LIME explainable AI technique outputs to interpret model focus areas, particularly on facial regions.

Strengths

• Model evaluation includes three distinct out-of-distribution experiments, demonstrating transferability.
• Performance metrics are highly detailed, including 99% accuracy, 1.0 precision, 0.97 recall, and 0.98 F1 score.
• External reliability was assessed on two additional public datasets: the Deepfake Dataset and the Deepfake Database.

Limitations

• The core dataset is extremely small at 5.5 KB, containing only model configurations and results, not the underlying image data.
• Geographic representation is heavily biased, with only 500 real images from Bangladesh combined with a larger external dataset of unspecified origin.
• Temporal coverage of the underlying image data is unknown, which may affect model relevance to newer deepfake generation techniques.

Provenance

Source

Figshare, authored by Md Shihab Reza.

Collection Method

Compiled from model training and evaluation experiments using a combined dataset of 500 real Bangladeshi images and a 140k real/fake faces dataset.

Freshness

Last updated in April 2026.

Geography

Includes facial images from Bangladesh; external dataset geography is unspecified.