ARNet

ARNet is a deep learning framework designed to address class imbalance in *video-based binary classification tasks, with a focus on *facial analysis and deepfake detection.
The framework introduces an adaptive resampling strategy that balances classes by controlling the number of frames extracted per video, while preserving all video sources.

---

📌 Key Features

•⁠ ⁠Adaptive Frame Resampling to handle class imbalance
•⁠ ⁠Video-Level Preservation (no video is discarded)
•⁠ ⁠CNN-Based Pipeline, compatible with AlexNet-style backbones
•⁠ ⁠Binary Classification Support
•⁠ ⁠Imbalance-Aware Evaluation Metrics (ROC, AUC, G-Mean)

---

🧠 Methodology

ARNet follows three main stages:

1.⁠ ⁠Frame Extraction
Frames are sampled from videos according to a configurable policy.

2.⁠ ⁠Adaptive Undersampling

• The class distribution is analyzed at the training stage.
• When imbalance is detected, the number of frames extracted from videos belonging to the majority class is reduced.
• Videos from the minority class retain a higher number of frames.
• This balances the dataset at the frame level while maintaining all video-level sources.

3.⁠ ⁠Model Training
Extracted frames are used to train a CNN-based binary classifier.

---

🏗️ Model Architecture

ARNet is architecture-agnostic. A typical configuration includes:

•⁠ ⁠Input: RGB facial frames
•⁠ ⁠Backbone: AlexNet (or similar CNN)
•⁠ ⁠Fully Connected Layers
•⁠ ⁠Output: Sigmoid activation for binary classification