Visual Attribution Limits and Graded Provenance Auditing in Commercial Video-Generation APIs
Organizations: Peking University · TokenRhythm Technologies
Abstract
Commercial video-generation APIs bill by model and tier, while intermediaries could substitute cheaper models, lower tiers, or cached clips. What can a delivered file establish about its origin? We combine measurements of passive attribution and cryptographic provenance into a graded audit. We test attribution probes and zero-shot detectors under re-encoding, downscaling, and frame-rate changes, and survey C2PA and GB 45438 metadata across 19 commercial endpoints. Only authenticated provenance supports accusations: tier contradictions or duplicate signed instances under fresh-generation contracts, excluding retries and idempotent returns. Unverifiable signed-endpoint deliveries trigger original-file demands under pass-through clauses; archive-reuse and delivery-parameter screens support review only. Six-model container and bitstream accuracy falls from 83.9% on raw files to 45.3% after uniform re-encoding, using different feature sets. Subsampling real references from 24 to 3 fps raises pooled D3 AUROC from 0.845 to 0.988. Only BytePlus authenticates tier claims under our pinned policy, including 187/187 audited deliveries across three Seedance tiers; tested remuxing and re-encoding leave processed Seedance files unverifiable. The audit identifies authenticated contradictions in 100/100 raw 720p signed tier-swap cases, with zero observed false flags on the tested honest pools.
Figures & tables
| Files | Feature set | Clips | Six-way acc. |
|---|---|---|---|
| Raw | All container and bitstream fields | 360 | 83.9% |
| Raw | Container and bitstream, provenance removed | 360 | 83.3% |
| N1 | Remuxed container and bitstream, no provenance | 360 | 77.5% |
| N2 | Bitstream statistics only (CRF-23 re-encode) | 360 | 45.3% |
| Pair (claimed sub.) | Claimed tuple | Substitute tuple | Diff. fields |
|---|---|---|---|
| Kling 3.0 Std Wan 3.0 | , 24 fps, 121f, no audio | , 30 fps, 150f, AAC 44.1 kHz, 2 ch, s | Every field |
| Seedance 2.5 HH 1.1 | , 24 fps, 121f, no audio | Same video; AAC 24 kHz, 2 ch, s | Audio track |
| Grok 1.5 1.5 Lite | , 24 fps, 121f, AAC 48 kHz, 2 ch, s | Identical except s | Offset only |
| Wan Prime Wan 3.0 | , 30 fps, 150f, AAC 44.1 kHz, 2 ch, s | Identical | None |
| Seedance 2.5 Mini/Fast | , 24 fps, 121f, no audio | Identical (720p) | None |
| Vendor: models | Regime | Identity field(s) | Granul. | Third-party verifiability |
|---|---|---|---|---|
| BytePlus: Seedance 2.5, 2.0 Fast, 2.0 Mini; 2.0 ( ) | C2PA | model_name , when , ID | Tier | Verified 187/187 across 3 tiers (tier = order; +1/1 on 2.0, ); GlobalSign R45 (Mozilla e-mail, not C2PA list); OCSP GOOD |
| Black Forest Labs: FLUX.3 video ( ) | C2PA | softwareAgent FLUX.3, ID | Model | Pinned-policy verified at model granularity; chains to GlobalSign R45 |
| Runway: Gen-4.5 ( ) | C2PA | agent + version 4.5.0 | Agent | Not verified: signingCredential.invalid despite C2PA-listed root |
| xAI: Grok Imagine 1.5, 1.5 Lite | C2PA | shared “Grok Imagine” | Product | Not authenticated under our pinned policy: 142/142 distinct per-clip self-signed CAs ( LOCAL USE ONLY ) |
| Alibaba: Wan 2.6, 2.7 ( each), 3.0, P; HH 1.1 | GB 45438 | shared producer, ProduceID | Vendor | Structure 164/164; buyer integrity untested (scheme/key not supplied); prefix matches 24/24 own & other tier |
| Kuaishou: Kling 3.0 Pro ( ), Std | GB 45438 | vendor ID, ProduceID , sm3 (prelim.) | Vendor | Structure 42/42; buyer integrity untested (scheme/key not supplied); 2nd session added self-signed C2PA ( ) |
Data Provenance
Data provenance tracks where data and AI outputs come from, how they are transformed, and what evidence supports their attribution and use.
See all 205 papers