Atomic-UACI: Inference-Time AI Content Provenance & Watermarking
⚠️ Status: Specification Stage (v2.1) — This repository contains the formal technical specification and whitepaper for the UACI Framework. Implementation is in active development. Benchmarks reflect design targets validated at specification stage.
The UACI Framework™ (Universal AI Content Identification System) is the first framework to atomically bind a cryptographic identifier, invisible multimodal watermark, and encrypted provenance vault during inference — across text, image, audio, video, and code.
Most existing approaches handle provenance after generation — post-hoc manifests, separate watermark specs, or siloed per-media solutions. UACI addresses this at the inference layer, before output is released.
| Gap | Current State |
|---|---|
| No standardised content identifiers | Each platform defines provenance differently |
| Watermarks stripped post-generation | No atomic binding at source |
| Opaque model pipelines | No inference-time auditability |
| Privacy exposure in provenance metadata | Plaintext manifests vulnerable |
┌─────────────────────────────────────────────────┐
│ UACI Framework │
├──────────┬──────────────┬────────────────────────┤
│ UACI │ StealthResist│ PAIR Router │
│ Protocol │ Watermark │ (inference middleware) │
├──────────┴──────────────┴────────────────────────┤
│ OpenVeri Toolkit (verification) │
├──────────────────────────────────────────────────┤
│ P3 Capsule (encrypted vault) │
└──────────────────────────────────────────────────┘
| Module | Function |
|---|---|
| UACI Protocol | Cryptographic globally unique identifier — version, timestamp, model fingerprint, content hash, CRC32 |
| StealthResist™ | Invisible multimodal watermark — DCT embedding for images, STFT for audio, token insertion for text |
| PAIR Router | Inference-time middleware — atomic injection of ID, watermark, and encrypted vault in a single pass |
| OpenVeri™ Toolkit | Open verification engine — extract watermark, validate header, decrypt vault, verify HMAC |
| P3 Capsule | Privacy-preserving encrypted container — AES-GCM provenance vault with GDPR erasure path |
| Prior Art | Limitation | UACI Difference |
|---|---|---|
| Google SynthID (US 11,706,256) | No encrypted vault; separate per-media specs | Unified atomic binding across all modalities |
| C2PA Spec v2.1 | Post-hoc cleartext manifests | Inference-time atomic binding |
| Intel GUID (US 11,514,365) | No watermark; post-generation only | Full cryptographic + watermark + vault pipeline |
| Operation | Median | 95th Percentile | Setup |
|---|---|---|---|
| ID Generation | 3 ± 0.5 ms | 6 ± 1.2 ms | Intel i7-1165G7, 16GB RAM |
| Watermark Embed | 42 ± 5 ms | 68 ± 8 ms | CPU only, COCO 1080p |
| Verification | 55 ± 6 ms | 80 ± 10 ms | CPU/GPU NVIDIA T4 |
| Throughput (batch 4) | 250 req/s | — | 4-core VM |
⚠️ Image watermark benchmarks validated at specification stage. Audio and video recovery metrics are roadmap items — see roadmap.
| Module | NIST SP 800-53 | ISO/IEC 42001 | EU AI Act / GDPR |
|---|---|---|---|
| UACI Protocol | SC-16, AU-10 | A.5, A.8 | Art 52 transparency |
| StealthResist | SC-7 | A.9 | — |
| P3 Capsule | MP-6, AC-17 | A.6 | GDPR Art 17 erasure |
atomic-uaci/
├── README.md
├── LICENSE
├── WHITEPAPER.md — Full technical whitepaper v2.1
├── docs/
│ ├── architecture.md — Five-module architecture detail
│ ├── security.md — STRIDE threat model and mitigations
│ ├── compliance.md — Governance and regulatory mapping
│ ├── benchmarks.md — Performance specifications
│ └── roadmap.md — Development roadmap
├── schemas/
│ └── p3-capsule.json — P3 Capsule provenance schema
└── specs/
└── uaci-protocol.md — UACI Protocol field specification
| Milestone | Target | Status |
|---|---|---|
| Specification v2.1 | Q1 2026 | ✅ Complete |
| POC Implementation | Q2 2026 | 🔄 In progress |
| Audio/Video Benchmarks | Q3 2026 | ⏳ Planned |
| Blockchain Registry Pilot | Q4 2026 | ⏳ Planned |
| ISO/IEC 42001 Engagement | Q1 2027 | ⏳ Planned |
| C2PA & W3C DID WG Contribution | Q1 2027 | ⏳ Planned |
MIT License — see LICENSE
Roshan George Thomas Independent AI Researcher — Bahrain ORCID: 0009-0002-1175-7749 Organisation: XwhyZ-WHYLD