Update README.md

README.md

@@ -22,18 +22,18 @@ The game uses a deep learning model built with **MobileNetV3Small** and transfer
 
 - **Base Model**: MobileNetV3Small (pre-trained on ImageNet)
 - **Architecture**: 
-  - Frozen MobileNetV3Small backbone for feature extraction
+  - MobileNetV3Small backbone for feature extraction (Unfreeze last 4 layers)
   - Global average pooling
   - Batch normalization + Dropout (0.15)
   - Dense layer (128 units) with ReLU activation
   - Batch normalization + Dropout (0.25)
   - Output layer with sigmoid activation for binary classification
 - **Data Augmentation**: Random rotation, translation, zoom, and horizontal flip
-- **Training**: 25 epochs with Adam optimizer (5e-4 learning rate)
+- **Training**: 150 epochs with Adam optimizer (1e-4 learning rate)
 - **Input Size**: 128×128 RGB images
 - **Task**: Binary classification (Real vs AI-generated faces)
 
-The model leverages transfer learning by using ImageNet pre-trained weights while keeping the base model frozen,
+The model leverages transfer learning by using ImageNet pre-trained weights and unfreeze last 4 layers,
 allowing it to effectively distinguish between authentic and AI-generated facial images.
 
 **Requirements**