DeepFake Analysis and Predictor

Overview:

A complete end-to-end pipeline for detecting DeepFakes using:

• Convolutional Neural Networks (CNNs) for face image artifacts

• Frequency-domain analysis (FFT + spectrograms) for synthetic voice patterns

• Head motion vector extraction for physical consistency detection

• Robust evaluation under adversarial perturbations (noise, compression, blur)

• A multimodal fusion classifier combining all signals

Project Directory Structure

deepfake_predictor:

├── data:

├── raw_videos: # Input .mp4 videos (real/fake)

├── extracted_frames: # Cropped face images

├── audio: # Extracted audio (.wav)

├── spectrograms: # Spectrogram images (.png)

├── labels.csv # id,label,motion format

├── scripts:

├── extract_and_label.py # Preprocess and generate labels

├── train.py # Train multimodal detector

├── eval.py # Evaluate on clean + adversarial distortions

├── inference.py # Predict real/fake on new video

├── models:

├── vision/vision_model.py

├── audio/audio_model.py

├── fusion/fusion_model.py

├── utils:

├── video_utils.py

├── audio_utils.py

├── motion_utils.py

├── requirements.txt

├── README.md