| """ |
| Explainability utilities for deepfake detection. |
| |
| Provides structured risk explanations with evidence summaries |
| for image, video, audio, and multimodal analysis. |
| """ |
|
|
|
|
| def explain_risk(score, model_scores=None): |
| """ |
| Generate structured risk explanation for image/video analysis. |
| |
| Args: |
| score: Final risk score (0.0 - 1.0). |
| model_scores: Optional dict of per-model scores for evidence. |
| |
| Returns: |
| str with detailed risk explanation. |
| """ |
| if score >= 0.60: |
| level = "AI-GENERATED" |
| desc = "Strong indicators of AI generation or manipulation" |
| else: |
| level = "AUTHENTIC" |
| desc = "No significant manipulation indicators" |
|
|
| explanation = f"{level} — {desc}" |
|
|
| if model_scores: |
| evidence = [] |
| if model_scores.get("vit_prob", 0) > 0.6: |
| evidence.append("ViT detected deepfake patterns") |
| if model_scores.get("face_prob", 0) > 0.6: |
| evidence.append("facial manipulation artifacts found") |
| if model_scores.get("forensic_prob", 0) > 0.5: |
| evidence.append("noise/ELA inconsistency detected") |
| if model_scores.get("frequency_prob", 0) > 0.5: |
| evidence.append("frequency-domain anomalies") |
| if model_scores.get("eff_prob", 0) > 0.6: |
| evidence.append("EfficientNet flagged AI generation") |
| if model_scores.get("dino_prob", 0) > 0.6: |
| evidence.append("DINOv2 detected synthetic features") |
| if model_scores.get("clip_prob", 0) > 0.6: |
| evidence.append("CLIP vision-language model detected manipulation signatures") |
|
|
| if evidence: |
| explanation += f". Evidence: {'; '.join(evidence)}" |
|
|
| return explanation |
|
|
|
|
| def explain_audio_risk(fake_prob): |
| """Explain audio deepfake risk level.""" |
| if fake_prob >= 0.60: |
| return "AI-GENERATED — AI-generated speech detected (voice cloning / TTS)" |
| else: |
| return "AUTHENTIC — Audio appears authentic" |
|
|
|
|
| def explain_multimodal(modality_scores, final_score): |
| """ |
| Generate explanation for multimodal fusion result. |
| |
| Args: |
| modality_scores: dict with image/video/audio scores. |
| final_score: fused risk score. |
| |
| Returns: |
| str explanation. |
| """ |
| active = {k: v for k, v in modality_scores.items() if v is not None} |
|
|
| if not active: |
| return "No modalities analyzed" |
|
|
| parts = [] |
| for mod, score in active.items(): |
| if score >= 60: |
| parts.append(f"{mod}: AI-generated ({score}%)") |
| else: |
| parts.append(f"{mod}: authentic ({score}%)") |
|
|
| if final_score >= 0.60: |
| verdict = "Strong evidence of manipulation across modalities" |
| else: |
| verdict = "Content appears authentic across analyzed modalities" |
|
|
| return f"{verdict}. Per-modality: {'; '.join(parts)}" |
|
|
