Model Performance Metrics
Evaluasi komprehensif EfficientNetB0 — Split 60/20/20 pada test set 6,000 gambar
95.30%
Test Accuracy
0.9902
AUC Score
0.95
F1-Score Fake
0.95
F1-Score Real
Dataset Information
Total Images30,000
Real Images15,000
Fake Images15,000
SourceKaggle — xhlulu
Train / Val / Test60 / 20 / 20 %
Test Samples6,000 gambar
Optimal Threshold0.67
Model Architecture
Base Model
EfficientNetB0 (ImageNet weights)
Transfer Learning
Phase 1: Base frozen (lr=3e-4) → Phase 2: Fine-tuning (lr=1e-5)
Custom Head
Dense(256) → Dropout(0.4) → Dense(128) → Dropout(0.3) → Sigmoid
Model File
realvsfake_effnetB0_final.h5
Confusion Matrix — Threshold Optimal 0.67 (Test Set: 6,000 gambar)
Predicted
Fake
Fake
Predicted
Real
Real
Actual: Fake
2,887True Negative · 96.2%
113False Positive · 3.8%
Actual: Real
169False Negative · 5.6%
2,831True Positive · 94.4%
Prediksi benar (5,718 / 6,000)
Prediksi salah (282 / 6,000)
Accuracy — Phase 1 (ep 1–20) & Phase 2 Fine-tuning (ep 21–60)
Loss — Phase 1 (ep 1–20) & Phase 2 Fine-tuning (ep 21–60)
ROC Curve
0.9902
AUC Score
95.30%
Akurasi (Threshold 0.67)
0.95
Sensitivity (Recall)