Upload agents/metadata_agent.py with huggingface_hub

agents/metadata_agent.py

@@ -44,6 +44,10 @@ def d02_software_check(img):
     return {"test":"Software Detection","score":s,"note":n}
 
 def d03_ela(img, quality=90):
+    # P4: Detect source format — ELA is only meaningful for JPEG inputs
+    source_format = getattr(img, 'format', None)
+    is_jpeg = source_format and source_format.upper() in ("JPEG", "JPG")
+    
     buf=io.BytesIO(); img_rgb=img.convert("RGB"); img_rgb.save(buf,"JPEG",quality=quality); buf.seek(0)
     resaved=Image.open(buf).convert("RGB"); ela=ImageChops.difference(img_rgb,resaved)
     ext=ela.getextrema(); mx=max(e[1] for e in ext) or 1
@@ -53,9 +57,12 @@ def d03_ela(img, quality=90):
     for i in range(0,h-bs,bs):
         for j in range(0,w-bs,bs): bm.append(float(np.mean(ea[i:i+bs,j:j+bs])))
     bm=np.array(bm); bstd=float(np.std(bm)); br=float(np.max(bm)-np.min(bm))
-    if bstd>8 and br>30: s,n=0.6,f"High ELA variance (σ={bstd:.1f}) — manipulation"
+    if not is_jpeg and float(np.std(ea))<1:
+        s,n=0.0,"PNG/lossless source — ELA comparison is lossless→JPEG, not meaningful"
+    elif bstd>8 and br>30: s,n=0.6,f"High ELA variance (σ={bstd:.1f}) — manipulation"
     elif bstd>4: s,n=0.3,f"Moderate ELA (σ={bstd:.1f})"
-    elif float(np.std(ea))<1: s,n=0.2,"Uniform ELA — AI"
+    elif float(np.std(ea))<1 and is_jpeg: s,n=0.2,"Uniform ELA on JPEG — possible AI"
+    elif float(np.std(ea))<1: s,n=0.0,"Uniform ELA (non-JPEG source)"
     else: s,n=-0.2,f"Consistent ELA (σ={bstd:.1f})"
     return {"test":"Error Level Analysis","block_std":round(bstd,3),"score":s,"note":n,"ela_image":ela_vis}
 
@@ -129,12 +136,36 @@ def d10_gps_plausibility(img):
     gps=exif.get(34853)
     if not gps: return {"test":"GPS Plausibility","score":0.0,"note":"No GPS data"}
     try:
-        # Check if GPS coordinates are physically possible
-        lat_ref=gps.get(1,"N"); lon_ref=gps.get(3,"E")
-        lat=gps.get(2,(0,0,0)); lon=gps.get(4,(0,0,0))
-        # Simple validation: lat ∈ [-90,90], lon ∈ [-180,180]
-        s,n=-0.2,f"GPS present ({lat_ref}, {lon_ref})"
-    except: s,n=0.0,"GPS parse error"
+        def _dms_to_dd(dms):
+            """Convert (degrees, minutes, seconds) tuple to decimal degrees."""
+            if isinstance(dms, (list, tuple)) and len(dms) == 3:
+                d = float(dms[0]) if not hasattr(dms[0], 'numerator') else float(dms[0])
+                m = float(dms[1]) if not hasattr(dms[1], 'numerator') else float(dms[1])
+                s_val = float(dms[2]) if not hasattr(dms[2], 'numerator') else float(dms[2])
+                return d + m/60.0 + s_val/3600.0
+            return float(dms)
+        
+        lat_ref = gps.get(1, "N")
+        lon_ref = gps.get(3, "E")
+        lat_raw = gps.get(2, (0,0,0))
+        lon_raw = gps.get(4, (0,0,0))
+        
+        lat = _dms_to_dd(lat_raw)
+        lon = _dms_to_dd(lon_raw)
+        if lat_ref == "S": lat = -lat
+        if lon_ref == "W": lon = -lon
+        
+        # Bounds check
+        if not (-90 <= lat <= 90) or not (-180 <= lon <= 180):
+            s,n = 0.4, f"Impossible GPS coordinates (lat={lat:.4f}, lon={lon:.4f})"
+        elif abs(lat) < 0.001 and abs(lon) < 0.001:
+            s,n = 0.3, f"GPS at Null Island (0,0) — likely fabricated"
+        elif abs(lat) < 0.1 and abs(lon) < 0.1:
+            s,n = 0.2, f"GPS near (0,0) — suspicious (lat={lat:.4f}, lon={lon:.4f})"
+        else:
+            s,n = -0.2, f"Plausible GPS ({lat:.4f}°{'N' if lat>=0 else 'S'}, {lon:.4f}°{'E' if lon>=0 else 'W'})"
+    except Exception as e:
+        s,n = 0.0, f"GPS parse error: {str(e)[:50]}"
     return {"test":"GPS Plausibility","score":s,"note":n}
 
 def d11_maker_note(img):
@@ -163,17 +194,11 @@ ALL_TESTS=[d01_exif_completeness,d02_software_check,d03_ela,d04_ai_metadata,d05_
            d10_gps_plausibility,d11_maker_note,d12_file_structure]
 
 def run_metadata_agent(img):
-    findings,scores=[],[]
-    ela_img=None
-    for fn in ALL_TESTS:
-        try:
-            r=fn(img); findings.append(r); scores.append(r["score"])
-            if "ela_image" in r: ela_img=r.pop("ela_image")
-        except Exception as e: findings.append({"test":fn.__name__,"error":str(e),"score":0})
-    avg=float(np.mean(scores)) if scores else 0.0; conf=min(1.0,0.5+0.5*abs(avg))
-    viol=[f["test"] for f in findings if f.get("score",0)>0.2]
-    comp=[f["test"] for f in findings if f.get("score",0)<-0.1]
-    rat=f"Metadata violations: {', '.join(viol)}." if viol else f"Metadata consistent: {', '.join(comp)}." if comp else "Metadata inconclusive."
-    for f in findings:
-        if f.get("note"): rat+=f" [{f['test']}]: {f['note']}."
-    return AgentEvidence("Metadata Agent",np.clip(avg,-1,1),conf,max(0,1-len(scores)/len(ALL_TESTS)),rat,findings,ela_img)
+    from agents.utils import run_agent_tests
+    findings_raw, avg, conf, fail, rat = run_agent_tests(ALL_TESTS, img, "Metadata Agent")
+    # Extract ELA image for visual evidence
+    ela_img = None
+    for f in findings_raw:
+        if "ela_image" in f:
+            ela_img = f.pop("ela_image")
+    return AgentEvidence("Metadata Agent", np.clip(avg,-1,1), conf, fail, rat, findings_raw, ela_img)