Model Performance Metrics for Deepfake Detection

A 5.5 KB Excel file containing classification report metrics for deepfake detection models. The dataset was published by Md Shihab Reza in April 2026 to accompany a study on the ARC-Net model. It compares the performance of ARC-Net against four other pre-trained deep learning models.

Use Cases

• Compare model accuracy, precision, recall, and F1-score across different deepfake detection architectures.
• Analyze the statistical significance of performance differences between classifiers using McNemar's test results.
• Benchmark the ARC-Net model's 99% accuracy and 1.0 precision against traditional and state-of-the-art methods.
• Evaluate the impact on false positive rate and probability calibration from incorporating Bangladeshi image data into training.

Strengths

• Provides direct performance metrics (accuracy, precision, recall, F1-score) for five distinct models.
• Includes results from three out-of-distribution experiments, including evaluation on South Asian images.
• Model evaluation was validated on external datasets (Deepfake Dataset and Deepfake Database) for generalizability.

Limitations

• The dataset is extremely small at 5.5 KB, suggesting it contains only summary statistics, not the underlying image or feature data.
• The exact row/column structure and specific metrics beyond the reported ones (e.g., per-class breakdowns) are unknown.
• The underlying training data of 500 real Bangladeshi images combined with 140k other faces is not included.

Provenance

Source

Md Shihab Reza, associated research study.

Collection Method

Generated from model evaluation experiments comparing ARC-Net against four pre-trained deep learning models.

Freshness

Last updated in April 2026.

Geography

Includes specific evaluation on images from Bangladesh and South Asia.