Deepfake Model Performance Metrics and Evaluation Results

ARC-Net achieved 99% accuracy and 0.98 F1 score in detecting deepfake images. Results compare five deep learning models, including ARC-Net and four pre-trained models, evaluated on a dataset containing 500 real images from Bangladesh and 140,000 real and fake faces. The evaluation was conducted by Md Shihab Reza and published in April 2026.

Use Cases

• Compare model performance metrics like accuracy, precision, recall, and F1-score across five different deep learning architectures.
• Analyze the impact of including Bangladeshi face images on false positive rates and probability calibration for models evaluated on South Asian data.
• Review results from three out-of-distribution experiments to assess model transferability across different deepfake datasets.
• Examine statistical significance between competing classifiers using results from McNemar's test reported in the evaluation.

Strengths

• Model evaluation includes results from five distinct deep learning models for direct comparison.
• Performance metrics are reported with high specificity, including 99% accuracy and 0.98 F1 score for the best model.
• External reliability was tested on two additional benchmark datasets (Deepfake Dataset and Deepfake Database).

Limitations

• The underlying evaluation dataset is small, built from only 500 real images from a single country.
• Dataset scope is limited to facial deepfakes, not covering other forms of synthetic media like audio or full-body video.
• The 5.5 KB file size suggests the dataset contains only summary statistics, not the raw image or feature data.

Provenance

Source

Md Shihab Reza, via figshare.

Collection Method

Model training and testing results from a research study introducing the ARC-Net architecture.

Freshness

Last updated in April 2026.

Geography

Includes specific evaluation on images from Bangladesh and South Asia.