Deploy Deepfake Shield Backend v2.0

detector.py

@@ -140,7 +140,11 @@ class FaceDeepfakeDetector:
             # Jalankan model untuk klasifikasi pada crop wajah dengan kedua model
             df_score = self._classify_image(face_crop, model_loader.face_deepfake_model)
             ai_score = self._classify_image(face_crop, model_loader.face_ai_edit_model)
-            score = min(df_score, ai_score)
+            
+            # Wvolf model (Deepfake) sangat akurat untuk wajah.
+            # dima806 (AI Edit) sering false-positive di foto DSLR/Webcam.
+            # Beri bobot 85% untuk Deepfake dan 15% untuk AI Edit agar akurat.
+            score = (df_score * 0.85) + (ai_score * 0.15)
             
             face_scores.append(score)
             face_boxes.append([x1, y1, x2 - x1, y2 - y1])
@@ -150,15 +154,23 @@ class FaceDeepfakeDetector:
         # lebih mudah dideteksi palsu daripada crop wajah.
         df_full_score = self._classify_image(frame_rgb, model_loader.face_deepfake_model)
         ai_full_score = self._classify_image(frame_rgb, model_loader.face_ai_edit_model)
-        full_frame_score = min(df_full_score, ai_full_score)
         
-        # --- Sistem Penilaian Agresif ---
+        # Untuk full frame, bobot diseimbangkan sedikit lebih tinggi untuk AI Edit (25%)
+        # karena konteks background berguna untuk deteksi AI generatif.
+        full_frame_score = (df_full_score * 0.75) + (ai_full_score * 0.25)
+        
+        # --- Sistem Penilaian Komprehensif ---
         # Kumpulkan semua score (Full Frame + Wajah)
         all_scores = [full_frame_score] + face_scores
         
-        # Ambil nilai probabilitas REAL terkecil (artinya paling FAKE).
-        # Jika salah satu komponen (background/wajah) terdeteksi palsu, anggap palsu.
-        representative_score = float(min(all_scores))
+        # Ambil rata-rata dari semua skor, namun condong ke skor yang paling mencurigakan
+        # Ini lebih seimbang daripada sekadar mengambil min() yang sering false-positive.
+        if len(all_scores) > 0:
+            avg_score = sum(all_scores) / len(all_scores)
+            min_score = min(all_scores)
+            representative_score = (avg_score * 0.6) + (min_score * 0.4)
+        else:
+            representative_score = full_frame_score
         
         logger.info(
             f"📊 Skor: Full Frame={full_frame_score:.3f}, Wajah={[round(s,3) for s in face_scores]} "

model_loader.py

@@ -77,7 +77,7 @@ class ModelLoader:
         self._log_device_info()
 
         # Muat model klasifikasi wajah
-        self._face_deepfake_model = self._load_model_pipeline("prithivMLmods/deepfake-detector-model-v1", "Deepfake")
+        self._face_deepfake_model = self._load_model_pipeline("Wvolf/ViT_Deepfake_Detection", "Deepfake")
         self._face_ai_edit_model = self._load_model_pipeline("dima806/deepfake_vs_real_image_detection", "AI Edit")
 
         # Muat model klasifikasi audio (CNN sederhana)