Deep Image Authentication

Multi-stage AI analysis combined with crowd-sourced visual review and expert sign-off for high-stakes image verification.

The Deep Image Authentication pipeline is designed for situations where a fast AI check is not sufficient. It combines forensic pre-screening and AI visual analysis with a five-person crowd review and a two-expert sign-off to produce a court-defensible verdict. Use this pipeline for images in legal proceedings, high-impact newsroom decisions, or compliance reporting where a single-layer automated result carries unacceptable risk. Because crowd and expert stages introduce human latency, this pipeline takes longer than Tier 1 alternatives — the trade-off is a substantially richer evidence trail and a consensus-gated result that carries the weight of multiple independent reviewers.

Pipeline Details

Property	Value
Pipeline ID	`verify-image-deep`
Tier	2
Estimated Duration	~80 min
Credit Cost	4,900 credits per run
Crowd Stages	Yes
Files Required	Yes — enriched image

Required Inputs

Files

Submit one enriched image file. Enrichment must be complete before starting this pipeline — the run will be rejected if the file has not been processed. Accepted formats: JPEG, PNG, WebP, TIFF, HEIC, AVIF.

Context Keys

Key	Required	Description
`claimed_source`	Optional	Reported origin or publisher of the image (e.g. `"Getty Images"`, `"Reuters"`)
`claimed_date`	Optional	Date the image is claimed to have been taken, ISO 8601 format (e.g. `"2024-07-14"`)
`claimed_location`	Optional	Geographic location where the image is claimed to have been captured (e.g. `"Paris, France"`)

Providing context keys significantly improves the quality of AI and crowd assessments. Workers are shown these claims and asked to evaluate their plausibility against visible image content.

Stages

Forensic Pre-Screen (AI)

The pipeline opens with a broad forensic sweep of the image file. The AI scans for common manipulation artefacts including EXIF inconsistencies, quantisation table mismatches, double-compression signatures, colour-space anomalies, and irregular noise patterns in high-frequency regions. Findings from this stage are surfaced to crowd workers in later stages as structured priming context, ensuring human reviewers focus their attention on the highest-risk areas.

Synthetic Content Screening (AI)

A dedicated AI stage scores the likelihood that the image is fully or partially AI-generated. It looks for content-credential (C2PA) tags, generator-tool signatures, and the absence of camera-specific EXIF fields expected from a genuine photograph. The stage writes a syntheticLikelihood score (0–100) and a list of syntheticIndicators to its artifacts — see Synthetic Data Detection.

Visual Analysis (AI)

A deeper AI inspection of the image's visual content layer. The model evaluates object consistency, shadow direction and length relative to visible light sources, perspective geometry, and the plausibility of spatial relationships between elements. Facial regions (if present) are assessed for blending boundaries and skin-texture regularity. This stage produces a structured set of visual anomaly scores that feed into evidence synthesis.

Crowd Context Review (Crowd)

Five crowd workers holding general and image_forensics skill tags independently review the image alongside the AI forensic findings and any supplied context claims. Each worker completes a structured SurveyJS questionnaire covering perceived authenticity, consistency of claimed metadata with visible content, and any anomalies the worker identifies independently. The stage gates on 60% consensus (three of five workers in agreement) before proceeding. The median time in this stage is approximately 45 minutes.

Evidence Synthesis (AI)

The AI aggregates findings from both forensic stages and all crowd worker responses into a unified evidence summary. Conflicting crowd responses are flagged explicitly. The synthesis stage produces a structured intermediate report that expert reviewers receive alongside the original image.

Expert Review (Expert Crowd)

Two specialist crowd workers holding the image_forensics expert tag independently review the image and the complete evidence summary produced in the previous stage. Each expert provides a structured verdict and written rationale. This stage requires 100% consensus — both experts must agree before the pipeline advances. This is the highest-confidence gate in the pipeline and the primary basis for court-admissible or compliance-grade verdicts.

Final Scorecard (AI Synthesis)

The pipeline closes with a weighted aggregation of all AI scores, crowd consensus signals, and expert verdicts into a single final verdict (authentic, manipulated, synthetic, or inconclusive) with a calibrated confidence score. The output includes a full audit scorecard listing every stage result, worker consensus breakdowns, and the complete evidence chain — suitable for inclusion in legal or editorial documentation.

Starting a Run

curl -X POST "https://api.crowdee.ai/v2/projects/{projectId}/verification-runs" \
  -H "X-API-Key: crw_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "pipeline_id": "verify-image-deep",
    "file_ids": ["file_01j9x..."],
    "context": {
      "claimed_source": "Getty Images",
      "claimed_date": "2024-07-14",
      "claimed_location": "Paris, France"
    }
  }'

Example Response

{
  "id": "run_01j9x...",
  "status": "completed",
  "verdict": "authentic",
  "confidence": 81,
  "pipeline_id": "verify-image-deep"
}

This pipeline takes approximately 80 minutes due to crowd stage latency. Poll GET /v2/projects/{projectId}/verification-runs/{runId} every 30–60 seconds, or implement a webhook for completion notification.

For a fast metadata-only check, see verify-image-metadata. If the Tier 1 result is inconclusive or the stakes are high, escalate to this pipeline on the same file.