Module 4 Lesson 5: Data provenance and integrity

If you don't know where your data came from, you can't trust what your AI says. Data Provenance (Lineage) and Integrity are the twin pillars of a secure data pipeline.

graph LR
    subgraph "Trust Chain"
    S[Source: Vendor/Sensor] -- "Sign & Hash" --> C[Ingestion: Verify Signature]
    C -- "Store as Immutable" --> D[Vault: WORM Storage]
    D -- "Verify Hash" --> T[Training Environment]
    T -- "Log Metadata" --> A[Lineage Audit Log]
    end

1. What is Data Provenance?

Provenance is the Accountability Trail for data.

• "This image was captured by Camera A on 2024-01-01, stored in S3 Bucket B, and processed by Script C before being used for Training."
• If you find a "Poisoned" sample, provenance tells you who uploaded it and when, allowing you to clear out other "Dirty" data from the same source.

---

2. Ensuring Data Integrity

Integrity ensures that the data wasn't changed after it was collected.

• Technique: Hashing (Checksums).
  • As soon as a dataset is finalized, calculate its SHA-256 hash.
  • Before training begins, re-calculate the hash. If they don't match, someone (or a piece of malware) has tampered with your data.

---

3. The "Chain of Custody" for ML

1. Source Verification: Use digital signatures to verify that data from a vendor is actually from that vendor.
2. Immutability: Store finalized training sets in "Write Once, Read Many" (WORM) storage.
3. Audit Logging: Every time a data scientist "views" or "edits" the training data, it should be recorded in a tamper-proof log (like CloudTrail or a dedicated security log).

---

4. Modern Tools for Lineage

• DVC (Data Version Control): Like Git but for large data files.
• MLflow: Tracks entire experiments, including which data led to which version of the model.
• OpenLineage: A standard for describing data movement between different systems (Spark, Airflow, SQL).

---

Exercise: The Investigator's Log

1. You discover a "Backdoor" trigger in your model. You have the hash of the dataset used for training. What is your first step?
2. Why is "Data Integrity" more difficult in a "Streaming" environment where the model is constantly learning from new user inputs?
3. What is the danger of using "Anonymous" data from a community hub without a signed manifest?
4. Research: What is "Software Bill of Materials" (SBOM) for data? (Often called a "Data Bill of Materials" or DBOM).

---

Summary

You have completed Module 4: Data Security and Data Poisoning. You now understand that data is a high-value security asset, how it can be poisoned with backdoors, why it might leak secrets, and how to build a "Trustworthy Pipeline" using provenance and integrity checks.

Next Module: The Silent Steal: Module 5: Model-Level Attacks.