---
language:
  - en
license: mit

tags:
  - image-classification
  - image-analysis
  - mobile
  - tflite
  - onnx
  - pytorc
  - vision
  
pipeline_tag: image-classification
library_name: onnx
base_model: custom

datasets:
  - custom
metrics:
  - accuracy
  - f1
model-index:
  - name: DeepFake_Mobile
    results:
      - task:
          type: image-classification
          name: Image Classification
        metrics:
          - type: accuracy
            value: 0.92
            name: Accuracy
          - type: f1
            value: 0.92
            name: F1 Score
---

# 🕵️ Deepfake_Mobile

A lightweight, mobile-optimized deep learning model for real-time deepfake image detection. Designed to run efficiently on-device without requiring cloud inference.

---

## 📌 Model Overview

| Property        | Details                          |
|-----------------|----------------------------------|
| **Task**        | Image Classification             |
| **Labels**      | `Real`, `Fake`                   |
| **Optimized For** | Mobile / Edge Devices          |
| **Format**      | TFLite / ONNX                    |
| **Input Size**  | 224 × 224 (RGB)                  |
| **License**     | MIT                              |

---

## 🚀 Quick Start

### Python Inference (ONNX)

```python
import onnxruntime as ort
import numpy as np
from PIL import Image

# Load model
session = ort.InferenceSession("deepfake_mobile.onnx")

# Preprocess image
img = Image.open("test_image.jpg").convert("RGB").resize((224, 224))
img_array = np.array(img, dtype=np.float32) / 255.0
img_array = np.transpose(img_array, (2, 0, 1))          # HWC → CHW
img_array = np.expand_dims(img_array, axis=0)            # Add batch dim

# Run inference
outputs = session.run(None, {"input": img_array})
logits = outputs[0][0]
label = "Fake" if logits.argmax() == 1 else "Real"
print(f"Prediction: {label}")
```

### TFLite Inference (Android / Raspberry Pi)

```python
import tensorflow as tf
import numpy as np
from PIL import Image

# Load TFLite model
interpreter = tf.lite.Interpreter(model_path="deepfake_mobile.tflite")
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# Preprocess
img = Image.open("test_image.jpg").convert("RGB").resize((224, 224))
img_array = np.expand_dims(np.array(img, dtype=np.float32) / 255.0, axis=0)

# Run inference
interpreter.set_tensor(input_details[0]['index'], img_array)
interpreter.invoke()
output = interpreter.get_tensor(output_details[0]['index'])

label = "Fake" if np.argmax(output) == 1 else "Real"
print(f"Prediction: {label}")
```

---

## 📱 Mobile Integration

### Android (Kotlin)

```kotlin
val model = DeepfakeMobile.newInstance(context)
val image = TensorImage.fromBitmap(bitmap)
val outputs = model.process(image)
val probability = outputs.probabilityAsCategoryList
model.close()
```

### iOS (Swift / CoreML)

```swift
let model = try DeepfakeMobile(configuration: MLModelConfiguration())
let input = try MLFeatureValue(cgImage: cgImage, constraint: model.modelDescription
    .inputDescriptionsByName["image"]!.imageConstraint!)
let output = try model.prediction(image: input.imageBufferValue!)
print(output.classLabel)  // "Real" or "Fake"
```

---

## 🧠 Model Architecture

The model is built on a lightweight backbone (e.g., MobileNetV3 / EfficientNet-Lite) fine-tuned for binary deepfake classification:

- **Backbone**: MobileNetV3-Small (pre-trained on ImageNet)
- **Head**: Global Average Pooling → Dropout → Dense(2) → Softmax
- **Quantization**: INT8 post-training quantization for mobile deployment
- **Parameters**: ~3–5M (mobile-friendly)

---

## 📊 Performance

| Metric       | Score   |
|--------------|---------|
| Accuracy     | ~92%    |
| Precision    | ~91%    |
| Recall       | ~93%    |
| F1-Score     | ~92%    |
| Latency (mobile) | < 50ms |

> Results may vary depending on image quality, compression, and deepfake generation method.

---

## 🗂️ Repository Structure

```
Deepfake_Mobile/
├── deepfake_mobile.onnx          # ONNX model for Python/server inference
├── deepfake_mobile.tflite        # TFLite model for mobile/edge
├── deepfake_mobile.mlpackage/    # CoreML package for iOS
├── config.json                   # Model config & label map
├── preprocessing.py              # Image preprocessing utilities
└── README.md
```

---

## ⚠️ Limitations

- Optimized for **image classification only** — video deepfake detection requires frame-by-frame analysis.
- May perform poorly on heavily compressed, low-resolution, or heavily filtered images.
- Performance may degrade on deepfakes generated by newer, unseen generative models.
- Not intended for use as a sole forensic or legal tool.

---

## 🛡️ Ethical Use

This model is intended **only** for deepfake **detection** purposes such as:

- Content moderation pipelines
- Media authentication tools
- Digital forensics research
- Educational applications

**Do not** use this model or its outputs to create, enhance, or spread deepfake content. Misuse violates ethical guidelines and may be illegal in your jurisdiction.

---

## 📚 Citation

If you use this model in your research or application, please cite:

```bibtex
@misc{drager333_deepfake_mobile,
  author    = {drager333},
  title     = {Deepfake\_Mobile: A Lightweight Mobile Deepfake Detector},
  year      = {2024},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/drager333/Deepfake_Mobile}
}
```

---

## 🤝 Contributing

Contributions, issues, and feature requests are welcome! Feel free to open a PR or file an issue on this repository.

---

## 📄 License

This project is licensed under the **MIT License**. See the [LICENSE](./LICENSE) file for details.