Full Video Authentication

Comprehensive video deepfake and manipulation detection combining AI forensics with crowd visual review and expert sign-off.

The Full Video Authentication pipeline is Crowdee's most thorough video verification offering. It sequentially applies technical container forensics, deepfake frame detection, audio authenticity analysis, crowd visual review, and a two-expert sign-off to produce a comprehensive, auditable verdict. Designed for investigative journalism, legal evidence review, and regulatory risk assessment, this pipeline is the appropriate choice when a video's authenticity is genuinely contested and the downstream consequences of an incorrect verdict are significant. The multi-layer approach means that no single analysis point — AI or human — determines the outcome alone; every stage's findings are carried forward and surfaced to subsequent reviewers.

Pipeline Details

Property	Value
Pipeline ID	`verify-video-full`
Tier	2
Estimated Duration	~120 min
Credit Cost	8,900 credits per run
Crowd Stages	Yes
Files Required	Yes — enriched video with extracted frames and audio track

Required Inputs

Files

Submit one enriched video file. Enrichment must be complete and must have produced extracted frames and a separate audio track before this pipeline can start — the run will be rejected if frame extraction or audio separation has not been completed. Accepted formats: MP4, MOV, MKV, WebM, AVI.

Context Keys

Key	Required	Description
`claimed_source`	Optional	Reported origin or publisher of the video (e.g. `"Ukrainian Parliament press pool"`)
`claimed_date`	Optional	Date the video is claimed to have been recorded, ISO 8601 format
`claimed_location`	Optional	Geographic location where the video is claimed to have been filmed
`claimed_event`	Optional	Described event or context for the video content (e.g. `"Parliamentary session 2024-03-12"`)

Providing event and source context is particularly valuable in this pipeline: crowd workers reviewing extracted frames are asked to assess whether visible content is consistent with the claimed time, place, and circumstances.

Stages

Technical Forensics (AI)

The pipeline opens by examining the video container's structural metadata: codec parameters, encoding history, bitrate profiles, timestamp sequences, and container-level anomalies. Re-encoding signatures, discontinuous GOP structures, and mismatched creation timestamps are catalogued and scored. These low-level technical signals are carried forward to all subsequent stages as foundational evidence.

Deepfake Screen (AI)

Extracted frames containing facial regions are evaluated for deepfake synthesis signatures. The AI assesses blending boundaries around face edges, unnatural skin texture regularity, temporal inconsistency between frames (flickering, blurring artefacts at hairlines and ears), and geometric implausibilities in facial landmarks. Results are expressed as per-frame anomaly scores and an aggregate deepfake probability for the video as a whole. This stage also writes a syntheticLikelihood score (0–100) and a list of syntheticIndicators to its artifacts — see Synthetic Data Detection.

Audio Authenticity (AI)

The separated audio track is analysed for synthetic speech indicators and audio-visual synchronisation issues. The AI evaluates vocoder fingerprints, prosodic unnaturalness, spectral discontinuities consistent with splicing, and lip-sync alignment across sampled frames. Desynchronisation events and confidence-scored anomaly windows are included in the findings passed to crowd workers.

Crowd Visual Review (Crowd)

Five crowd workers holding general and video_forensics skill tags independently review a curated set of extracted frames alongside the AI findings from all three prior stages. Workers complete a structured assessment questionnaire covering their perception of authenticity, consistency of claimed context with visible content, and any independently observed anomalies. The stage gates on 60% consensus (three of five workers in agreement) before proceeding. This is typically the longest stage, averaging 60–70 minutes of elapsed time.

Evidence Aggregation (AI)

All technical signals and crowd worker responses are aggregated into a single structured evidence document. Conflicting worker assessments are explicitly flagged and weighted. The aggregated evidence is formatted for expert review in the following stage, with the highest-risk findings surfaced prominently.

Expert Review (Expert Crowd)

Two specialist crowd workers holding the video_forensics expert tag independently review the original video, the extracted frames, and the complete aggregated evidence document. Each expert provides a structured verdict and written rationale covering technical and perceptual dimensions. This stage requires 100% consensus — both experts must reach the same verdict before the pipeline advances. Disagreement triggers a manual escalation flag in the run record.

Final Scorecard (AI Synthesis)

A weighted aggregation of all AI scores, crowd consensus signals, and expert verdicts produces the final run verdict (authentic, manipulated, synthetic, or inconclusive) with a calibrated confidence score. The output includes a full audit report with stage-by-stage findings, worker consensus breakdowns, frame-level anomaly references, and a complete evidence chain suitable for legal, editorial, or compliance 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-video-full",
    "file_ids": ["file_01j9x..."],
    "context": {
      "claimed_source": "Ukrainian Parliament press pool",
      "claimed_event": "Parliamentary session 2024-03-12"
    }
  }'

Example Response

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

Estimated duration is ~120 minutes. This is Crowdee's most comprehensive pipeline. Plan for asynchronous handling — do not hold an HTTP connection open. Poll GET /v2/projects/{projectId}/verification-runs/{runId} every 30–60 seconds, or use webhooks for completion notification.

For faster video checks, see verify-video-technical. Ensure the video is fully enriched with extracted frames and a separated audio track before starting this pipeline — enrichment status is validated at run creation time.