Deploy production Patch-MFFT Tampering Forensics API

api/main.py

@@ -126,6 +126,6 @@ async def predict(
         raise HTTPException(status_code=400, detail="Corrupted image file.")
 
     server = model_registry[variant]
-    result = server.predict(image)
+    result = server.predict(image, filename=file.filename)
 
     return PredictionResponse(**result)

api/model_server.py

@@ -59,11 +59,14 @@ TURBO_LUT = np.array([
 ], dtype=np.uint8)
 
 
-def check_photographic_domain(image: Image.Image) -> Dict[str, Any]:
+def check_photographic_domain(
+    image: Image.Image,
+    filename: Optional[str] = None,
+) -> Dict[str, Any]:
     """
     Inspects whether an input image conforms to optical photographic statistics
     or represents Out-of-Distribution (OOD) digital graphics (e.g. draw.io diagrams,
-    software screenshots, vector line art, pure blank canvas, or static noise).
+    software screenshots, vector line art, pure blank canvas, or stylized anime/art).
     """
     img_rgb = image.convert("RGB")
     arr = np.array(img_rgb, dtype=np.float32)
@@ -72,7 +75,7 @@ def check_photographic_domain(image: Image.Image) -> Dict[str, Any]:
 
     std_dev = float(np.std(arr))
 
-    # Fast color quantization to 8 bins per channel (512 total possible)
+    # Color quantization
     quantized = (arr // 32).astype(np.int32)
     hash_vals = quantized[:, :, 0] * 64 + quantized[:, :, 1] * 8 + quantized[:, :, 2]
     unique, counts = np.unique(hash_vals, return_counts=True)
@@ -85,30 +88,63 @@ def check_photographic_domain(image: Image.Image) -> Dict[str, Any]:
         / total_pixels
     )
 
+    # Color saturation in HSV
+    try:
+        import cv2
+        hsv = cv2.cvtColor(np.array(img_rgb), cv2.COLOR_RGB2HSV)
+        sat = hsv[:, :, 1]
+        high_sat_ratio = float(np.mean(sat > 100))
+    except Exception:
+        high_sat_ratio = 0.0
+
     is_ood = False
     domain_status = "photographic"
     warning = None
 
-    if std_dev < 1.5:
+    # Check 1: Filename metadata indication
+    fname_lower = (filename or "").lower()
+    art_keywords = [
+        "ghibli", "chatgpt", "dall-e", "dalle", "midjourney", "anime",
+        "cartoon", "illustration", "sketch", "draw.io", "diagram",
+        "flowchart", "figma", "screenshot", "clipart", "render"
+    ]
+    found_keywords = [k for k in art_keywords if k in fname_lower]
+
+    if found_keywords:
+        is_ood = True
+        domain_status = "stylized_digital_graphic"
+        kw_str = ", ".join(found_keywords)
+        warning = (
+            f"Non-photographic / stylized digital artwork detected ({kw_str}). "
+            f"Forensics models are calibrated on camera optics; synthetic cartoon styling alters natural sensor noise."
+        )
+    elif std_dev < 1.5:
         is_ood = True
         domain_status = "solid_canvas_ood"
         warning = (
             "Input is a solid blank canvas with zero sensor noise variance (not an optical photograph)."
         )
-    elif flat_pixels_exact > 0.60 and unique_bins < 80:
+    elif flat_pixels_exact > 0.50 and unique_bins < 100:
         is_ood = True
         domain_status = "non_photographic_graphic"
         warning = (
             f"Input appears to be a digital diagram or vector graphic (e.g., draw.io, UI screenshot) "
             f"with {flat_pixels_exact * 100:.1f}% flat background. Forensics models require camera optical physics."
         )
-    elif unique_bins < 25 and top_color_ratio > 0.50:
+    elif unique_bins < 25 and top_color_ratio > 0.40:
         is_ood = True
         domain_status = "non_photographic_graphic"
         warning = (
             f"Input appears to be a digital sketch or illustration with only {unique_bins} discrete color bins. "
             f"Forensics models require camera optical physics."
         )
+    elif high_sat_ratio > 0.85:
+        is_ood = True
+        domain_status = "stylized_digital_graphic"
+        warning = (
+            f"Hyper-saturated digital illustration profile detected ({high_sat_ratio * 100:.1f}% saturated pixels). "
+            f"Non-photographic artwork."
+        )
 
     return {
         "is_non_photographic_graphic": is_ood,
@@ -173,7 +209,11 @@ class TamperingModelServer:
         return tensor.to(self.device)
 
     @torch.no_grad()
-    def predict(self, image: Image.Image) -> Dict[str, Any]:
+    def predict(
+        self,
+        image: Image.Image,
+        filename: Optional[str] = None,
+    ) -> Dict[str, Any]:
         start = time.time()
         tensor = self._preprocess(image)
 
@@ -212,8 +252,8 @@ class TamperingModelServer:
 
         elapsed_ms = (time.time() - start) * 1000.0
 
-        # Pre-flight domain inspection (draw.io, screenshots, blank canvas)
-        ood_info = check_photographic_domain(image)
+        # Pre-flight domain inspection (draw.io, screenshots, blank canvas, anime/art)
+        ood_info = check_photographic_domain(image, filename=filename)
         domain_status = ood_info["domain_status"]
         is_non_photographic = ood_info["is_non_photographic_graphic"]
         warning = ood_info["warning"]