🚀 Visual Inference - Deepfake Detection

🔍 Visual Inference is a deepfake detection system that utilizes Xception-based feature extraction and Transformer-based sequence modeling to classify real vs. fake images. This project provides a Dockerized solution to simplify model deployment and inference.

---

📌 Features

✔️ Xception-based feature extraction for frame analysis
✔️ Transformer Encoder for sequence modeling
✔️ Cross-Attention Mechanism to refine embeddings
✔️ Docker support for easy deployment
✔️ Inference statistics to analyze model performance

---

🔧 Installation & Setup

1️⃣ Clone the Repository

git clone git@github.com:gocenalper/visual-inference.git
cd visual-inference

2️⃣ Install Dependencies (For Local Usage)

pip install torch torchvision timm tqdm pillow

---

🐳 Running with Docker

We provide a lightweight Docker image for running the model without manually installing dependencies.

1️⃣ Build the Docker Image

docker build -t dfdc-inference .

2️⃣ Run the Docker Container (Mounting Dataset & Code)

Run the following command to mount your dataset and code inside the container:

docker run --rm -it -v "$(pwd)":/app dfdc-inference

3️⃣ Run with GPU Support (Optional)

If your machine has CUDA-enabled GPUs, use:

docker run --gpus all --rm -it -v "$(pwd)":/app dfdc-inference

---

🖼 Dataset Structure

The dataset should be mounted in the following format:

/DFDC/
    ├── REAL/
    │   ├── TRAIN/
    │   │   ├── video_0001/
    │   │   │   ├── frame_01.jpg
    │   │   │   ├── frame_02.jpg
    │   │   ├── video_0002/
    │   ├── TEST/
    │   ├── VAL/
    ├── FAKE/
    │   ├── TRAIN/
    │   │   ├── video_0003/
    │   │   │   ├── frame_01.jpg
    │   │   │   ├── frame_02.jpg
    │   ├── TEST/
    │   ├── VAL/

---

🚀 Running Inference

Once the Docker container is running, the model will process test images and print real-time statistics:

🔍 Running Untrained Model Inference on Test Data...

📌 Image 1: Predicted = FAKE, Actual = REAL, ❌ Incorrect
📌 Image 2: Predicted = REAL, Actual = FAKE, ❌ Incorrect
📌 Image 3: Predicted = REAL, Actual = REAL, ✅ Correct
📌 Image 4: Predicted = FAKE, Actual = FAKE, ✅ Correct

📊 **Inference Statistics (Before Training)**
🔹 Total Images Processed: 1000
🟢 Real Predictions: 500 (50.0%)
🔴 Fake Predictions: 500 (50.0%)
✅ Correct Predictions: 495 (49.5%) Accuracy

---