Une API pour détecter les deepfakes dans les images et vidéos.

{
  "schemaVersion": "deepfakepolicy.report.v1",
  "reportId": "DFP-SAMPLE-IMAGE-0001",
  "generatedAt": "2026-09-13T12:00:00.000Z",
  "source": {
    "modality": "image",
    "sha256": "55c824c6118dfb18ef5156c3e7c6b5bf4cf9cd53ef44c3f1b5f5efa27c45910e",
    "name": "demo-portrait.jpg",
    "mediaType": "image/jpeg",
    "format": "JPG",
    "url": null,
    "sizeBytes": 63578,
    "pageCount": null,
    "characterCount": null
  },
  "analysis": {
    "mode": "quick",
    "status": "complete",
    "creditsUsed": 1,
    "method": "DeepfakePolicy evidence synthesis"
  },
  "evidencePath": [
    {
      "id": "provenance",
      "label": "Provenance",
      "status": "limited",
      "summary": "No verified origin credential or complete chain of custody was available for this report.",
      "signals": []
    },
    {
      "id": "metadata",
      "label": "Metadata",
      "status": "reviewed",
      "summary": "Available format, coverage and metadata observations were reviewed.",
      "signals": [
        {
          "label": "Submitted format",
          "finding": "JPG",
          "score": null,
          "scoreLabel": null
        }
      ]
    },
    {
      "id": "media-signals",
      "label": "Media signals",
      "status": "not-available",
      "summary": "No separate human-readable media observation was available for this modality.",
      "signals": []
    },
    {
      "id": "model-signals",
      "label": "Model signals",
      "status": "reviewed",
      "summary": "The returned classifications are listed separately, including unavailable checks and any segment findings.",
      "signals": [
        {
          "label": "AI-generated image",
          "finding": "Indicators found",
          "score": 82,
          "scoreLabel": "82% signal score"
        },
        {
          "label": "Deepfake face manipulation",
          "finding": "No strong signal in this example",
          "score": 6,
          "scoreLabel": "6% signal score"
        }
      ]
    },
    {
      "id": "web-evidence",
      "label": "Public web evidence",
      "status": "not-available",
      "summary": "Automated public-source research was not included in this analysis.",
      "signals": []
    }
  ],
  "verdict": {
    "code": "suspicious",
    "label": "Suspicious signals detected",
    "aiLikelihood": 82,
    "confidenceBand": "not-rated",
    "summary": "The analysed media contains signals that require review. The individual findings identify what to examine.",
    "rationale": [
      "The AI-generation detector returned an elevated signal.",
      "No independently verified capture history is supplied in this example."
    ]
  },
  "recommendedActions": [
    "Preserve the earliest available, highest-quality original and its surrounding context.",
    "Confirm the publisher, identities and important claims through an independent channel.",
    "Compare key frames, audio segments or images with an attributable original source.",
    "Escalate high-stakes use to a qualified forensic reviewer before acting."
  ],
  "limitations": [
    "This is a probabilistic assessment, not proof of authenticity, authorship or intent.",
    "Compression, editing, transcription, translation, short samples and unseen generation methods can change results.",
    "A low score does not certify authenticity, and a high score does not identify a creator.",
    "Public-source research can miss private, deleted, paywalled or unindexed material."
  ],
  "privacy": {
    "originalStoredWithReport": false,
    "researchContributionSeparate": true,
    "reportVisibility": "response-only",
    "statement": "The report contains normalized findings, not the submitted source. Any optional research contribution is governed and deleted separately."
  }
}

Cet exemple fictif d’image illustre une analyse standard. La recherche de sources est une option distincte ; les éléments disponibles dépendent du fichier et du mode choisi. aiLikelihood est un signal du modèle, pas une probabilité de fraude ni une garantie d’authenticité. Les noms des champs API restent identiques entre les intégrations.