metadata

language: en
tags:
  - deepfake
  - image-classification
  - computer-vision
  - swin-transformer
  - efficientnet
  - deep-learning
license: mit
datasets:
  - custom
metrics:
  - accuracy
model-index:
  - name: DeepFake Detection Swin-EfficientNet Hybrid
    results:
      - task:
          type: image-classification
          name: Image Classification
        dataset:
          name: Custom DeepFake Dataset
          type: images
        metrics:
          - type: accuracy
            value: 0.9867

🧠 Deepfake Detection using Swin Transformer

A robust deepfake detection model leveraging the microsoft/swin-base-patch4-window7-224 architecture. Built and trained on the "Hemg/deepfake-and-real-images" dataset with oversampling and image augmentations to ensure class balance and improved generalization.

📌 Table of Contents

📦 Cell1: Install Required Packages
📥 Cell2: Load Dataset & Apply Oversampling
🏷️ Cell3: Label Mapping + Split Dataset
🧠 Cell4: Load Swin Transformer + Transforms
📊 Cell5: Collator + Metrics + CSV/Plot Export
⚙️ Cell6: Training Arguments + Trainer Setup
🚀 Cell7: Train and Evaluate
📈 Cell8: Loss & Accuracy Plots

📦 Cell1: Install Required Packages

pip install -q datasets transformers evaluate imbalanced-learn
pip install -U transformers

📥 Cell2: Load Dataset & Apply Oversampling

Used the Hemg/deepfake-and-real-images dataset and applied RandomOverSampler from imblearn to balance the Real and Fake class distribution.

🏷️ Cell3: Label Mapping + Split Dataset

Mapped class labels to integers using ClassLabel and split the dataset using stratification into 60% training and 40% testing.

🧠 Cell4: Load Swin Transformer + Transforms

Loaded Swin Transformer from Hugging Face Transformers.
Defined different transforms for training and evaluation using torchvision.transforms.

📊 Cell5: Collator + Metrics + CSV/Plot Export

Used a custom collator to batch image tensors and labels.
Saved evaluation metrics (accuracy, precision, recall, f1) and confusion matrix as .csv and .png.