Upload basic_explainer.py

src/explainers/basic_explainer.py

@@ -128,14 +128,32 @@ class BasicExplainer:
                 supports_fake += 1
 
         conflict = (supports_fake > 0 and supports_real > 0)
+        
+        # -------------------- Suspiciously clean detection --------------------
+        # If ALL forensic cues are below threshold (supports_real > 0 and supports_fake == 0),
+        # AND the prediction is REAL, this could indicate a modern generator that evades detection.
+        # Flag as potentially suspicious if all cues are "clean" but confidence isn't very high.
+        suspiciously_clean = (supports_fake == 0 and supports_real >= 2 and 
+                              prediction_label == 0 and confidence < 0.98)
 
         # -------------------- Triage decision (narrative only) --------------------
         triage_label = base_label_str
         if self.enable_triage and conflict and confidence < self.triage_conf_threshold:
             triage_label = "UNCERTAIN"
+        elif self.enable_triage and suspiciously_clean and confidence < 0.95:
+            # Modern generators like Flux may evade all forensic cues
+            triage_label = "UNCERTAIN"
 
         # Intro sentence
-        if triage_label == "UNCERTAIN":
+        if triage_label == "UNCERTAIN" and suspiciously_clean:
+            explanation_parts.append(
+                f"⚠️ **CAUTION**: The detector predicts this image is **{base_label_str}** "
+                f"with {confidence_str} confidence ({confidence:.2f}), "
+                f"but ALL forensic cues are below threshold. This could indicate a modern generator "
+                f"(like Flux, DALL-E 3, or Midjourney v6) that evades traditional forensic detection. "
+                f"**Manual review recommended.**"
+            )
+        elif triage_label == "UNCERTAIN":
             explanation_parts.append(
                 f"The detector predicts this image is **{base_label_str}** "
                 f"with {confidence_str} confidence ({confidence:.2f}), "
@@ -170,12 +188,12 @@ class BasicExplainer:
                 if prediction_label == 0:
                     explanation_parts.append(
                         f"- **Noiseprint**: fingerprint lies within the range seen in training real images "
-                        f"(mismatch={nm:.2f} ≤ {thr_nm:.2f}), supporting the REAL hypothesis."
+                        f"(mismatch={nm:.2f} <= {thr_nm:.2f}), supporting the REAL hypothesis."
                     )
                 else:
                     explanation_parts.append(
                         f"- **Noiseprint**: fingerprint lies within the range seen in training real images "
-                        f"(mismatch={nm:.2f} ≤ {thr_nm:.2f}), but other forensic cues indicate synthesis."
+                        f"(mismatch={nm:.2f} <= {thr_nm:.2f}), but other forensic cues indicate synthesis."
                     )
             cues_used += 1
 
@@ -225,25 +243,59 @@ class BasicExplainer:
                         f"and conflicts with the REAL prediction."
                     )
                 cues_used += 1
-            # If fp ≤ thr_fp we treat it as weak / neutral, so we skip.
-
-        if cues_used == 0:
-            explanation_parts.append(
-                "No individual forensic cue strongly deviated from the training distribution; "
-                "the decision is based on a subtle combination of features."
-            )
+            elif prediction_label == 1:
+                # Even if below threshold, mention it if prediction is FAKE and it's close to threshold
+                if fp > thr_fp * 0.8:  # Within 80% of threshold
+                    explanation_parts.append(
+                        f"- **Frequency spectrum**: peakiness ({fp:.2f}) is moderately elevated "
+                        f"(threshold: {thr_fp:.2f}), contributing to the FAKE classification."
+                    )
+                    cues_used += 1
 
-        # -------------------- Data-driven drivers (optional) --------------------
+        # -------------------- Data-driven drivers (show what actually drove the decision) --------------------
         if contributions:
             sorted_contribs = sorted(contributions.items(), key=lambda x: abs(x[1]), reverse=True)
-            top = sorted_contribs[:top_k_contributions]
-            pos = [f"{name} ({val:+.2f})" for name, val in top if val > 0]
-            neg = [f"{name} ({val:+.2f})" for name, val in top if val < 0]
+            # Show top 5-8 features for better explanation
+            top = sorted_contribs[:max(top_k_contributions, 8)]
+            pos = [(name, val) for name, val in top if val > 0]
+            neg = [(name, val) for name, val in top if val < 0]
 
             if pos:
-                explanation_parts.append(f"- **Top pushes toward FAKE**: {', '.join(pos)}")
+                explanation_parts.append(f"\n**Features driving FAKE classification:**")
+                # Show top 5-8 features that push toward FAKE
+                pos_display = [f"{name} ({val:+.3f})" for name, val in pos[:8]]
+                explanation_parts.append(f"- {', '.join(pos_display)}")
             if neg:
-                explanation_parts.append(f"- **Top pushes toward REAL**: {', '.join(neg)}")
+                explanation_parts.append(f"\n**Features supporting REAL classification:**")
+                # Show top 3-5 features that push toward REAL
+                neg_display = [f"{name} ({val:+.3f})" for name, val in neg[:5]]
+                explanation_parts.append(f"- {', '.join(neg_display)}")
+        elif not contributions and (cues_used == 0 or (prediction_label == 1 and cues_used < 2)):
+            # If no strong forensic cues but high confidence, explain it's a combination
+            explanation_parts.append(
+                f"\n**Note**: While the primary forensic cues (Noiseprint, Residuals, FFT) don't individually "
+                f"strongly indicate synthesis, the model's decision is based on a combination of many features "
+                f"including DCT coefficients, FFT radial profiles, residual statistics, and other frequency-domain "
+                f"characteristics. The high confidence ({confidence:.1%}) suggests these subtle patterns collectively "
+                f"indicate synthetic generation."
+            )
+            
+            # List some of the other features that might be contributing
+            other_features = []
+            if 'dct_mean' in features:
+                other_features.append("DCT coefficients")
+            if 'fft_radial_mean' in features:
+                other_features.append("FFT radial profiles")
+            if 'residual_skew' in features:
+                other_features.append("residual statistics")
+            if 'residual_kurtosis' in features:
+                other_features.append("residual distribution shape")
+            
+            if other_features:
+                explanation_parts.append(
+                    f"The model analyzes {', '.join(other_features)} and other frequency-domain patterns "
+                    f"that collectively indicate synthetic generation, even when individual cues are subtle."
+                )
 
         # In high-conflict cases, add a final triage note
         if triage_label == "UNCERTAIN" and not is_ood: