update ALL

app.py

@@ -34,6 +34,31 @@ if os.environ.get("SPACE_ID"):
 # Available detectors based on launcher.py
 DETECTORS = ['ALL', 'R50_TF', 'R50_nodown', 'CLIP-D', 'P2G', 'NPR']
 
+# Performance weights from Table III (Non-shared) and Table IV (Social Media: FB, TL, X)
+# Weights are averaged between Non-shared and Social Media performance.
+# R50_TF (R50-E2P): 
+#   Real (TNR): (0.98 + 0.91) / 2 = 0.95
+#   Fake (TPR): (0.85 + 0.65) / 2 = 0.75
+# R50_nodown (R50-ND): 
+#   Real (TNR): (1.00 + 0.97) / 2 = 0.98
+#   Fake (TPR): (0.79 + 0.76) / 2 = 0.77
+# CLIP-D: 
+#   Real (TNR): (0.97 + 0.94) / 2 = 0.95
+#   Fake (TPR): (0.93 + 0.91) / 2 = 0.92
+# P2G: 
+#   Real (TNR): (1.00 + 1.00) / 2 = 1.00
+#   Fake (TPR): (0.91 + 0.56) / 2 = 0.74
+# NPR: 
+#   Real (TNR): (0.99 + 1.00) / 2 = 1.00
+#   Fake (TPR): (0.60 + 0.12) / 2 = 0.36
+DETECTOR_WEIGHTS = {
+    'R50_TF': {'real': 0.95, 'fake': 0.75},
+    'R50_nodown': {'real': 0.98, 'fake': 0.77},
+    'CLIP-D': {'real': 0.95, 'fake': 0.92},
+    'P2G': {'real': 1.00, 'fake': 0.74},
+    'NPR': {'real': 1.00, 'fake': 0.36}
+}
+
 def process_image(image_path):
     """
     Check if image is larger than 1024x1024 and central crop it if necessary.
@@ -136,8 +161,8 @@ def predict(image_path, detector_name):
             if not results:
                 return "Error: No results obtained from detectors.", None
             
-            votes_real = 0
-            votes_fake = 0
+            votes_real = 0.0
+            votes_fake = 0.0
             confidences = []
             labels = []
             colors = []
@@ -160,12 +185,14 @@ def predict(image_path, detector_name):
                 colors.append(color)
                 total_conf += score
                 
-                # Voting logic
+                # Weighted Voting logic
+                # Only count vote if confidence > 0.6
                 if score > 0.6:
+                    weights = DETECTOR_WEIGHTS.get(det, {'real': 1.0, 'fake': 1.0})
                     if pred == 'fake':
-                        votes_fake += 1
+                        votes_fake += weights['fake']
                     elif pred == 'real':
-                        votes_real += 1
+                        votes_real += weights['real']
             
             # Majority Voting
             if votes_real > votes_fake:
@@ -179,9 +206,9 @@ def predict(image_path, detector_name):
             
             # Explanation
             if verdict == "REAL":
-                explanation = f"Considering the results obtained by all models, the analyzed image results, with an average confidence of {avg_conf:.4f}, not produced by a generative AI."
+                explanation = f"Considering the results obtained by all models (weighted by their historical performance), the analyzed image results, with an average confidence of {avg_conf:.4f}, not produced by a generative AI."
             elif verdict == "FAKE":
-                explanation = f"Considering the results obtained by all models, the analyzed image results, with an average confidence of {avg_conf:.4f}, produced by a generative AI."
+                explanation = f"Considering the results obtained by all models (weighted by their historical performance), the analyzed image results, with an average confidence of {avg_conf:.4f}, produced by a generative AI."
             else:
                 explanation = f"The result is uncertain. The detectors produced unconsistent results. The average confidence is {avg_conf:.4f}."
             
@@ -268,7 +295,8 @@ with demo:
 
     ### Understanding the Results produced by "ALL"
     *   Runs all available detectors (R50_TF, R50_nodown, CLIP-D, P2G, NPR) sequentially on the input image.
-    *   Produces a **Majority Vote** verdict (Real/Fake) considering only confident predictions (> 0.6). Also generates a **Confidence Plot** visualizing each model's score and a textual **Explanation** of the consensus.
+    *   Produces a **Weighted Majority Vote** verdict (Real/Fake). Each model's vote is weighted by its historical performance (TPR for Fakes, TNR for Reals) on benchmark datasets. Only confident predictions (> 0.6) are counted.
+    *   Also generates a **Confidence Plot** visualizing each model's score and a textual **Explanation** of the consensus.
     *   In the plot, **Green** bars indicate a **Real** prediction, while **Red** bars indicate a **Fake** prediction.
 
 

verify_all_option.py

@@ -41,10 +41,11 @@ def test_all_option():
             
             if isinstance(fig, plt.Figure):
                 print("Figure created successfully.")
-                # Optional: check plot content if needed
+                if "weighted by their historical performance" in text:
+                 print("Text output confirms weighted voting is used.")
             else:
-                print("Figure creation failed or None returned.")
-                
+                 print("Text output MISSING weighted voting confirmation.")
+
             expected_verdict = "produced by a generative AI" # Majority Fake
             if expected_verdict in text:
                  print("Verdict seems correct (Fake majority).")