Upload agents/modality_detector.py with huggingface_hub

agents/modality_detector.py

@@ -124,17 +124,21 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     sharp_fraction = float(np.mean(sharp_region))
     blur_fraction = float(np.mean(blur_region))
     
+    # CRITICAL: Check absolute sharpness level, not just relative
+    # Real portrait photos have genuinely sharp foreground (textures, edges, pores)
+    # AI images are smooth everywhere — even the "sharp" region is soft
+    peak_sharpness = float(np.percentile(sharpness, 95))
+    median_sharpness = float(np.median(sharpness))
+    
     # Check if blur is very uniform (computational vs optical)
     blur_values = sharpness[blur_region] if np.any(blur_region) else np.array([0])
     blur_uniformity = 1.0 - min(float(np.std(blur_values)) / (float(np.mean(blur_values)) + 1e-9), 1.0)
     
-    # Check transition abruptness: in portrait mode the edge between sharp/blur is steep
-    # Compute gradient of sharpness map
+    # Check transition abruptness
     sharpness_grad = np.hypot(
         sobel(sharpness, axis=0),
         sobel(sharpness, axis=1)
     )
-    # High max gradient at the transition = abrupt = computational
     max_transition = float(np.percentile(sharpness_grad, 99))
     mean_transition = float(np.mean(sharpness_grad))
     transition_abruptness = max_transition / (mean_transition + 1e-9)
@@ -143,24 +147,33 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     indicators["blur_fraction"] = round(blur_fraction, 3)
     indicators["blur_uniformity"] = round(blur_uniformity, 3)
     indicators["transition_abruptness"] = round(transition_abruptness, 3)
+    indicators["peak_sharpness"] = round(peak_sharpness, 2)
+    indicators["median_sharpness"] = round(median_sharpness, 2)
+    
+    # Portrait mode detection — requires BOTH relative and absolute sharpness
+    # A real portrait photo has genuinely sharp foreground detail
+    # An AI-generated smooth image has low peak sharpness even if center > edges
+    has_genuine_detail = peak_sharpness > 10.0  # Real photos have Laplacian variance > 10
     
-    # Portrait mode detection — relaxed thresholds + multiple signals
     portrait_signals = 0
-    if blur_fraction > 0.2 and sharp_fraction > 0.1:
+    if blur_fraction > 0.2 and sharp_fraction > 0.1 and has_genuine_detail:
         portrait_signals += 1
-    if blur_uniformity > 0.5:
+    if blur_uniformity > 0.5 and has_genuine_detail:
         portrait_signals += 1
-    if transition_abruptness > 5:  # Abrupt edge = computational segmentation
+    if transition_abruptness > 5 and has_genuine_detail:
         portrait_signals += 1
     
-    # Strong portrait mode: at least 2 of 3 signals
-    if portrait_signals >= 2:
+    # Strong portrait mode: at least 2 of 3 signals AND genuine foreground detail
+    if portrait_signals >= 2 and has_genuine_detail:
         scores["PORTRAIT_MODE"] = scores.get("PORTRAIT_MODE", 0) + 0.3 * portrait_signals
         indicators["portrait_mode_signature"] = True
-    elif portrait_signals == 1 and blur_fraction > 0.25:
+    elif portrait_signals == 1 and blur_fraction > 0.25 and has_genuine_detail:
         scores["PORTRAIT_MODE"] = scores.get("PORTRAIT_MODE", 0) + 0.2
         indicators["portrait_mode_weak"] = True
     
+    if not has_genuine_detail:
+        indicators["low_detail_image"] = True  # Smooth everywhere = possible AI
+    
     # ── 4. Screenshot detection ───────────────────────────────────────
     # Screenshots have: perfect pixel edges, UI elements, uniform background areas
     edge_mag = np.hypot(sobel(gray, 0), sobel(gray, 1))
@@ -201,12 +214,18 @@ def detect_modality(img: Image.Image) -> ModalityResult:
         confidence = min(1.0, scores[modality])
     
     # Override: portrait mode wins over messaging/smartphone when detected
-    # (a portrait photo sent via WhatsApp should get PORTRAIT_MODE adjustments,
-    #  which are a superset of MESSAGING adjustments for the tests that matter)
     if scores.get("PORTRAIT_MODE", 0) > 0.2:
         modality = "PORTRAIT_MODE"
         confidence = min(1.0, scores["PORTRAIT_MODE"])
     
+    # SAFETY GUARD: If image has no genuine photographic detail (low peak sharpness),
+    # it's likely AI-generated, not a real camera photo. In this case, do NOT apply
+    # any modality suppression — let all forensic tests fire at full strength.
+    if indicators.get("low_detail_image", False):
+        modality = "UNKNOWN"
+        confidence = 0.2
+        indicators["modality_override"] = "Low-detail image — suppression disabled to avoid protecting AI content"
+    
     # ── 7. Build score adjustments ────────────────────────────────────
     # Merge adjustments when multiple modalities detected
     # (e.g., portrait mode photo sent via messaging app gets BOTH sets)
@@ -249,7 +268,7 @@ def _get_modality_adjustments(modality: str) -> dict:
             "Vignetting cos⁴θ": 0.3,          # Smartphones don't follow cos⁴θ
             "HF Noise Structure": 0.3,         # Blur region has different noise
             "Noise Spatial Frequency": 0.3,    # Same reason
-            "CFA Nyquist": 0.5,               # Computational processing removes CFA
+            "CFA Nyquist": 0.25,              # Computational processing destroys CFA traces entirely
             # Spectral tests affected by computational photography
             "Spectral Slope 1/f²": 0.5,       # Frequency-selective sharpening steepens slope
             "Spectral Symmetry": 0.4,          # Depth-based processing creates asymmetry