Upload thresholds_calibrated.json with huggingface_hub

thresholds_calibrated.json

@@ -1,61 +1,45 @@
 {
   "reconstruction_thresholds": {
-    "description": "Calibrated MSE reconstruction error thresholds",
-    "dataset_real": "CIFAR-10 test set",
-    "dataset_fake": "Random Noise test set",
-    "image_size": [
-      128,
-      128
-    ],
-    "normalization": "mean=0.5, std=0.5 (range [-1, 1])",
     "thresholds": {
       "strict": {
-        "value": 0.012767740525305271,
-        "description": "High confidence - 99th percentile of real images",
-        "expected_tpr": 1.0,
-        "expected_fpr": 0.01
+        "value": 0.055736735463142395,
+        "expected_fpr": 0.01,
+        "description": "99th percentile - very low false positive rate"
       },
       "balanced": {
-        "value": 0.009065581485629082,
-        "description": "Balanced - 95th percentile of real images",
-        "expected_tpr": 1.0,
-        "expected_fpr": 0.05
+        "value": 0.039442386478185654,
+        "expected_fpr": 0.05,
+        "description": "95th percentile - balanced detection"
       },
       "sensitive": {
-        "value": 0.00931915920227766,
-        "description": "Sensitive - mean + 2*std of real images",
-        "expected_tpr": 1.0,
-        "expected_fpr": 0.0448
-      },
-      "optimal": {
-        "value": 0.20403903722763062,
-        "description": "Optimal separation between distributions",
-        "expected_tpr": 1.0,
-        "expected_fpr": 0.0
+        "value": 0.03894944489002228,
+        "expected_fpr": 0.025,
+        "description": "Mean + 2 std - sensitive detection"
       }
     },
     "statistics": {
-      "real_images": {
-        "mean": 0.004293377045542002,
-        "std": 0.0025128910783678293,
-        "median": 0.003766049398109317,
-        "min": 0.00019204482669010758,
-        "max": 0.02400028333067894,
-        "percentile_95": 0.009065581485629082,
-        "percentile_99": 0.012767740525305271
-      },
-      "fake_images": {
-        "mean": 0.40168634057044983,
-        "median": 0.4016798138618469,
-        "min": 0.3955906331539154,
-        "max": 0.4081667363643646,
-        "percentile_5": 0.3990125060081482
-      },
-      "separation": {
-        "mean_difference": 0.39739295840263367,
-        "ratio": 93.55953216552734
-      }
-    },
-    "note": "These thresholds are calibrated on CIFAR-10 vs Random Noise. For real deepfake detection, recalibrate on actual fake images."
+      "real_mean": 0.0183913204818964,
+      "real_median": 0.015646658837795258,
+      "real_std": 0.010279062204062939,
+      "real_95th": 0.039442386478185654,
+      "real_99th": 0.055736735463142395,
+      "fake_mean": 0.3526953160762787,
+      "fake_median": 0.34715136885643005,
+      "separation": 19.177268981933594,
+      "num_real_samples": 28000,
+      "num_fake_samples": 10000
+    }
+  },
+  "model_info": {
+    "best_epoch": 1,
+    "best_val_loss": 0.010239500302954443,
+    "total_epochs": 2,
+    "datasets": [
+      "CIFAR-10",
+      "CIFAR-100",
+      "STL-10-train",
+      "STL-10-unlabeled"
+    ],
+    "training_time_minutes": 21.360870798428852
   }
 }