ARC-Net Model Parameters for Bangladeshi Deepfake Detection

Model parameters detail the architecture and performance metrics of ARC-Net, a deepfake detection model. The model was trained on a dataset combining 500 real images from Bangladesh with a 140k real and fake faces dataset. Md Shihab Reza authored the study, with parameters published on figshare in April 2026.

Use Cases

• Analyze attention and residual convolutional layer configurations for model replication.
• Compare ARC-Net's 99% accuracy and 1.0 precision metrics against other pre-trained models.
• Evaluate the impact of incorporating Bangladeshi real images on false positive rate reduction.
• Review Grad-CAM and LIME explainable AI results for model transparency.
• Assess McNemar's statistical test results for significance between competing classifiers.

Strengths

• Model achieved 99% accuracy, 1.0 precision, 0.97 recall, and 0.98 F1 score on test data.
• Training dataset included 140k real and fake faces combined with 500 specific Bangladeshi real images.
• External validation performed on Deepfake Dataset and Deepfake Database datasets.

Limitations

• The dataset of parameters is only 5.5 KB, indicating a very limited scope of information.
• The core training dataset's geographic coverage is heavily biased towards Bangladesh.
• The number of rows or detailed parameter records is unknown.

Provenance

Source

Md Shihab Reza via figshare.

Collection Method

Parameters derived from training and testing the ARC-Net model.

Freshness

Published April 2026.

Geography

Bangladesh, with cross-dataset validation.