A simple binary classifier to detect AI-generated videos using EfficientNet-V2-L and ConvNeXt-Base models with adaptive frame extraction.

1. Install dependencies:

pip install -r requirements.txt

2. Set up Hugging Face authentication (optional):

export HUGGINGFACE_TOKEN=your_token_here

Download the dataset and extract frames adaptively:

python frame_extractor.py --download --output ./data_frames

This will:

• Download the DeepAction dataset locally
• Extract 1-7 frames per video based on duration
• Save frames as JPG images in ./data_frames/

Train both models on the extracted frames:

python train.py

Test trained models and see confidence scores:

python evaluate.py

• EfficientNet-V2-L: 145M parameters, pretrained on ImageNet
• ConvNeXt-Base: 88M parameters, pretrained on ImageNet

Both models are modified for 2-class binary classification (Real vs AI-generated).

Uses the faridlab/deepaction_v1 dataset from Hugging Face, which contains:

• Real videos
• AI-generated videos

The system downloads videos locally and extracts frames adaptively:

• ≤1s: 1 frame
• ≤5s: 2 frames
• ≤10s: 3 frames
• ≤20s: 5 frames
• ≤30s: 6 frames

• 30s: 7 frames (cap)

• Extracted frames saved to data_frames/ directory
• Trained models saved to models/ directory (only classifier weights)
• Training logs saved to training.log
• Evaluation shows accuracy, precision, recall, F1-score, and confidence scores