Skip to content

project.md

Latest commit

 

History

History
166 lines (127 loc) · 5.46 KB

project.md

File metadata and controls

166 lines (127 loc) · 5.46 KB

Project Brief: Sensor-PKI Authenticity (SPKIA)

Public-Key Cryptography at the Image Sensor Level + C2PA-aware AI Media Detection + ML-enhanced Detection

1) Objective

Develop a privacy-preserving verification platform that determines whether an image or video was:

  • Captured by a real camera (via sensor-level cryptographic proof and/or valid C2PA credentials),
  • Likely AI-generated (Nano Banana, Sora, DALL·E, etc.), or
  • Provenance unknown.

The system must integrate cryptographic proofs and machine learning classifiers to ensure robustness against adversarial manipulation.

2) Scope

In-scope

  • Verification via C2PA manifests.
  • Verification via sensor-level PKI signatures.
  • ML/AI-based detection pipeline for AI media classification.
  • Privacy-first operation: ephemeral processing only, no permanent storage.

Out-of-scope

  • Persistent content storage or user profiling.

3) Architecture

Frontend (React)

  • Simple drag-and-drop or URL input.
  • Displays authenticity result and rationale.
  • Option to show cryptographic and ML-based verification details.

Backend (Python/FastAPI)

  • Orchestration of verification pipeline.
  • Calls C2PA validator, Sensor-PKI verifier, and ML classifiers.
  • Stateless; horizontally scalable.

Database (MongoDB)

  • Stores job states, extracted metadata, and ML features temporarily.
  • TTL expiration (default: 24h).

Deployment (Docker)

  • Containers: frontend, API, workers, Mongo, Nginx.
  • Portable and easily deployed to cloud or on-prem environments.

4) Cryptographic Layer

Sensor-PKI

  • Cameras sign capture hashes with secure hardware keys.
  • Public key infrastructure maintained by manufacturers.
  • Signatures embedded as COSE/CBOR envelopes or as C2PA assertions.

C2PA

  • Validate provenance claims, trust chains, edit history.
  • Expose results to users.

5) ML-Based Detection Layer

When cryptographic proofs are absent or stripped, use ML classifiers:

  • Convolutional Neural Networks (CNNs) for pixel-level artifact detection.
  • Noise Residual & PRNU-based models for identifying real vs synthetic sources.
  • Transformers (Vision + Multimodal) for AI-model artifact detection.
  • Metadata anomaly detection using ML to classify inconsistencies (e.g., frame cadence, EXIF tags).

Classifier Arms Race Strategy

  • Continuously retrain models against new AI generators.
  • Maintain benchmark sets of AI-generated vs. real content.
  • Ensemble classifiers (cryptographic + forensic + ML) for final decision.

6) Verification Pipeline

  1. Intake: File upload or URL provided.
  2. Hashing & Pre-checks.
  3. C2PA Verification (if present).
  4. Sensor-PKI Verification (if present).
  5. ML Detection Pipeline (if C2PA/PKI missing or inconclusive).
    • Extract noise features, encoder toolmarks, metadata.
    • Run ensemble classifiers.
  6. Decision Logic:
    • Authentic (Camera)
    • Authentic (C2PA)
    • Likely AI-generated
    • Unknown
  7. Return Results to frontend.
  8. Purge Artifacts after TTL expiry.

7) Data & Privacy

  • No raw content retention.
  • Ephemeral job storage in MongoDB.
  • Explicit “Delete Now” option for users.
  • Transparency page: clearly explains privacy guarantees.

8) API Contract

  • POST /verify — Upload file/URL. Returns { job_id }.
  • GET /verify/{job_id} — Returns status, authenticity label, cryptographic proofs, ML classifier results.
  • DELETE /verify/{job_id} — Force immediate deletion.
  • GET /health — Service health check.

9) Data Model (MongoDB)

  • jobs: { id, hash, status, label, reasons[], createdAt } (TTL).
  • proofs: { job_id, c2pa, sensor_pki, ml_results }.
  • metrics: anonymized aggregate counts (e.g., % AI, % authentic).

10) Deployment

  • Docker Compose for development.
  • Kubernetes-ready for production scaling.
  • Secrets managed via Vault/KMS.
  • TLS via Nginx reverse proxy.

11) KPIs

  • Precision/Recall on benchmark datasets.
  • Coverage: % with valid proofs (C2PA/PKI).
  • Detection performance: ML classifier F1-score > 90%.
  • Latency: <2s per image, <4s per short video.
  • User comprehension: clear differentiation between “Unknown” and “Likely AI”.

12) Implementation Timeline

  • Weeks 1–2: Requirements, trust anchor setup.
  • Weeks 3–6: C2PA verification implementation.
  • Weeks 5–8: Sensor-PKI signature verification.
  • Weeks 7–10: ML classifier integration (CNN + PRNU + metadata).
  • Weeks 9–12: Ensemble decision logic + testing.
  • Weeks 12–14: Security hardening, deployment, documentation.

13) Risks

  • Rapid evolution of AI image generators.
  • Partial adoption of C2PA by platforms.
  • Potential adversarial bypass of ML classifiers.
  • Governance of manufacturer trust anchors.

14) Documentation Deliverables

  • System Whitepaper (cryptography + ML).
  • Privacy & Security Policy.
  • API Reference.
  • ML Training Dataset Protocols.
  • Threat Model Analysis.

15) Why Hybrid Cryptography + ML Works

  • Positive proofs (C2PA, sensor-PKI) → strong cryptographic guarantees.
  • ML classifiers → resilience against provenance stripping and AI artifacts.
  • Ensemble decision → layered assurance, user-facing clarity.