Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- image-classification
|
| 5 |
+
- deepfake-detection
|
| 6 |
+
- tensorflow
|
| 7 |
+
- tflite
|
| 8 |
+
datasets:
|
| 9 |
+
- custom
|
| 10 |
+
metrics:
|
| 11 |
+
- accuracy
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# Deepfake Detector Model
|
| 15 |
+
|
| 16 |
+
Este modelo detecta si una imagen es REAL o FAKE (generada/manipulada).
|
| 17 |
+
|
| 18 |
+
## Modelo
|
| 19 |
+
|
| 20 |
+
- **Arquitectura:** ResNet50 con Transfer Learning
|
| 21 |
+
- **Framework:** TensorFlow / TensorFlow Lite
|
| 22 |
+
- **Input:** Imágenes RGB de 128x128 píxeles
|
| 23 |
+
- **Output:** Probabilidad sigmoid (0=FAKE, 1=REAL)
|
| 24 |
+
- **Threshold:** 0.5
|
| 25 |
+
|
| 26 |
+
## Uso
|
| 27 |
+
|
| 28 |
+
### Con TensorFlow Lite (Python)
|
| 29 |
+
|
| 30 |
+
```python
|
| 31 |
+
import tensorflow as tf
|
| 32 |
+
import numpy as np
|
| 33 |
+
from PIL import Image
|
| 34 |
+
from huggingface_hub import hf_hub_download
|
| 35 |
+
|
| 36 |
+
# Descargar modelo
|
| 37 |
+
model_path = hf_hub_download(
|
| 38 |
+
repo_id="juandaram/deepfake-detector",
|
| 39 |
+
filename="model.tflite"
|
| 40 |
+
)
|
| 41 |
+
|
| 42 |
+
# Cargar modelo
|
| 43 |
+
interpreter = tf.lite.Interpreter(model_path=model_path)
|
| 44 |
+
interpreter.allocate_tensors()
|
| 45 |
+
|
| 46 |
+
# Preparar imagen
|
| 47 |
+
image = Image.open("tu_imagen.jpg").convert('RGB')
|
| 48 |
+
image = image.resize((128, 128))
|
| 49 |
+
img_array = np.array(image, dtype=np.float32) / 255.0
|
| 50 |
+
img_batch = np.expand_dims(img_array, axis=0)
|
| 51 |
+
|
| 52 |
+
# Predecir
|
| 53 |
+
input_details = interpreter.get_input_details()
|
| 54 |
+
output_details = interpreter.get_output_details()
|
| 55 |
+
interpreter.set_tensor(input_details[0]['index'], img_batch)
|
| 56 |
+
interpreter.invoke()
|
| 57 |
+
output = interpreter.get_tensor(output_details[0]['index'])
|
| 58 |
+
|
| 59 |
+
prediction = "REAL" if output[0][0] > 0.5 else "FAKE"
|
| 60 |
+
confidence = output[0][0] if output[0][0] > 0.5 else (1 - output[0][0])
|
| 61 |
+
|
| 62 |
+
print(f"Prediction: {prediction}")
|
| 63 |
+
print(f"Confidence: {confidence:.3f}")
|
| 64 |
+
```
|
| 65 |
+
|
| 66 |
+
### Con SavedModel (Python)
|
| 67 |
+
|
| 68 |
+
```python
|
| 69 |
+
import tensorflow as tf
|
| 70 |
+
from huggingface_hub import snapshot_download
|
| 71 |
+
|
| 72 |
+
# Descargar modelo completo
|
| 73 |
+
model_dir = snapshot_download(repo_id="juandaram/deepfake-detector")
|
| 74 |
+
|
| 75 |
+
# Cargar modelo
|
| 76 |
+
model = tf.saved_model.load(f"{model_dir}/saved_model")
|
| 77 |
+
infer = model.signatures['serving_default']
|
| 78 |
+
|
| 79 |
+
# Usar igual que arriba...
|
| 80 |
+
```
|
| 81 |
+
|
| 82 |
+
## Métricas
|
| 83 |
+
|
| 84 |
+
- Validation Accuracy: ~84%
|
| 85 |
+
- Training Epochs: 5
|
| 86 |
+
|
| 87 |
+
## Clases
|
| 88 |
+
|
| 89 |
+
- 0: FAKE (imagen generada/manipulada)
|
| 90 |
+
- 1: REAL (imagen auténtica)
|
| 91 |
+
|
| 92 |
+
## Preprocesamiento
|
| 93 |
+
|
| 94 |
+
Las imágenes deben:
|
| 95 |
+
1. Convertirse a RGB
|
| 96 |
+
2. Redimensionarse a 128x128
|
| 97 |
+
3. Normalizarse dividiendo por 255.0 (rango [0, 1])
|
| 98 |
+
|
| 99 |
+
## Limitaciones
|
| 100 |
+
|
| 101 |
+
- El modelo puede tener sesgo hacia la clase FAKE
|
| 102 |
+
- Funciona mejor con imágenes similares al dataset de entrenamiento
|
| 103 |
+
- Requiere imágenes de buena calidad
|
| 104 |
+
|
| 105 |
+
## Licencia
|
| 106 |
+
|
| 107 |
+
MIT
|
| 108 |
+
|
| 109 |
+
## Contacto
|
| 110 |
+
|
| 111 |
+
Para preguntas o problemas, abre un issue en el repositorio.
|