MuhammadQasimShabbir3/deepfake-image-detectector

Deepfake and Real Images Detection

This repo accompanies the Kaggle notebook: https://www.kaggle.com/code/muhammadbehroz/deepfake-and-real-images-dectection/notebook

It trains an image classifier (EfficientNet-B7) to distinguish deepfake vs real images and records training/validation metrics and visualizations.

Highlights

• Model: EfficientNet-B7 (2-class)
• Best validation accuracy: 92.26% (10 epochs)
• Test accuracy: 92.36%
• F1 score: 0.9236
• AUC: 0.9748

Results

From results/outputs/run_enhanced_v1/test_results.json:

Metric	Value
Test loss	0.3284
Test accuracy	92.36%
F1 score	0.9236
AUC	0.9748
Precision	0.9086
Recall	0.9407
Test samples	10,905
Fake / Real	5,492 / 5,413

Confusion matrix (rows = true, cols = predicted):

[[4980,  512],
 [ 321, 5092]]

Repo layout

• deepfake-and-real-images-dectection.ipynb: primary workflow (data prep, training, evaluation)
• results/: training artifacts and visualizations
  • outputs/run_enhanced_v1/: metrics JSON, plots, and saved model

Reproducibility

Key settings and environment info are captured in:

• results/outputs/run_enhanced_v1/reproducibility_info.json

Notable settings:

• Seed: 42
• Model: efficientnet_b7
• Batch size: 4
• Learning rate: 1e-6

How to run

The notebook is the main entry point. You can:

1. Open the Kaggle notebook link above to run end-to-end in Kaggle.
2. Or open deepfake-and-real-images-dectection.ipynb locally in Jupyter.

Note: The notebook includes the full dependency setup and dataset source. If running locally, install the same packages shown in the notebook environment.

Acknowledgements

• Kaggle notebook: https://www.kaggle.com/code/muhammadbehroz/deepfake-and-real-images-dectection/notebook