Deepfake Detection Model .

A machine learning model that classifies media (image, video, or audio) as Real or Fake using metadata and signal-quality features rather than raw pixel/audio analysis.

Overview

This project trains a Logistic Regression classifier on a metadata dataset describing media samples (e.g. lip-sync consistency, visual artifacts, lighting inconsistencies, compression level, platform, and generation method) to predict whether the sample is real or AI-generated ("fake").

Dataset

The model expects a CSV file named deepfake_detection_metadata_dataset.csv with the following columns:

Column	Description
media_id	Unique identifier for the media sample
media_type	Type of media (Image, Video, Audio)
content_category	Content category (News, Interview, Social Media, etc.)
face_count	Number of faces detected in the media
audio_present	Whether audio is present (Yes/No)
lip_sync_score	Score indicating lip-sync consistency
visual_artifacts_score	Score indicating presence of visual artifacts
compression_level	Compression level of the media file
lighting_inconsistency_score	Score indicating lighting inconsistencies
source_platform	Platform the media was sourced from
generation_method	Method used to generate fake media (GAN, Diffusion, VoiceClone, etc.), if applicable
label	Target variable — Real or Fake

Pipeline

1. Load & explore data — inspect shape, head/tail, dtypes, and missing values.
2. Handle missing values — generation_method (missing for real media) is imputed with its mode.
3. Encode categorical features — one-hot encoding via pd.get_dummies() on media_type, content_category, audio_present, source_platform, and generation_method.
4. Train/test split — 80/20 split (random_state=42).
5. Train model — sklearn.linear_model.LogisticRegression.
6. Evaluate — confusion matrix and accuracy score on the held-out test set.
7. Visualize — box plots comparing lip_sync_score, visual_artifacts_score, and lighting_inconsistency_score between real and fake samples.
8. Save model — trained model is serialized to deepfake_model.pkl with joblib.

Requirements

pandas
numpy
matplotlib
seaborn
scikit-learn
joblib

Install with:

pip install pandas numpy matplotlib seaborn scikit-learn joblib

Usage

1. Place deepfake_detection_metadata_dataset.csv in the project directory.
2. Run the notebook (deepfake_model.ipynb) cell by cell, or export it to a script:

   jupyter nbconvert --to script deepfake_model.ipynb
   python deepfake_model.py

3. The trained model is saved as deepfake_model.pkl, and box plot figures are saved as <feature>_boxplot.png.

Loading the saved model

import joblib

model = joblib.load("deepfake_model.pkl")
predictions = model.predict(X_new)

Results

The model achieves high accuracy on the test split (confusion matrix and accuracy score are printed in the notebook).

Note: Accuracy came out at 100% on this dataset, which is unusually high for a real-world classifier. This is very likely because generation_method is only missing for Real samples (and filled in for Fake samples), so its one-hot encoding leaks the label. Before using this model on new data, consider dropping or re-deriving generation_method so the model relies on genuine signal-quality features (lip_sync_score, visual_artifacts_score, lighting_inconsistency_score, compression_level) instead.

Limitations & Future Work

• Trained on metadata/scores rather than raw media (no image, video frame, or audio waveform analysis).
• Potential label leakage via generation_method (see note above) should be addressed before deployment.
• No hyperparameter tuning, cross-validation, or regularization scaling was applied (LogisticRegression also raised a convergence warning — consider scaling features or increasing max_iter).
• Could be extended with a CNN/RNN pipeline on raw video frames or audio spectrograms for true content-based deepfake detection.

License

Add a license of