Data Poisoning Detection: Why Your Fine-Tuning Pipeline Needs Provenance Controls

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ethics-safety-data-poisoning-detection-adults

1. Which cryptographic technique is most appropriate for verifying training data provenance throughout a fine-tuning pipeline?

   1. End-to-end encryption of data stores
   2. Digital signatures on model weights
   3. OAuth tokens for API authentication
   4. Cryptographic hashes of individual data samples

2. Statistical anomaly detection on training data typically monitors which of the following?

   1. Label distribution, feature distribution, and outlier presence
   2. GPU memory utilization during training
   3. User engagement metrics with deployed models
   4. Model inference latency patterns

3. Why must a trusted evaluation set remain completely separate from the training pipeline?

   1. To comply with GDPR data minimization requirements
   2. To prevent data leakage that would inflate performance metrics
   3. To ensure the model sees only fresh data during testing
   4. To reduce computational costs during evaluation

4. What limitation prevents AI systems from detecting all forms of data poisoning?

   1. AI systems lack sufficient computational power for large-scale analysis
   2. AI cannot process unstructured data like text and images
   3. AI requires labeled data to identify poisoning patterns
   4. AI cannot detect poisoning that perfectly mimics legitimate data distribution

5. Which audit area examines vendor attestations and internal access controls for training data?

   1. Statistical anomaly detection
   2. Trigger-pattern evaluation
   3. Evaluation set integrity
   4. Supply-chain trust

6. After training, behavioral evaluation against suspected trigger patterns serves which purpose?

   1. To compress model size for deployment
   2. To measure model accuracy on standard benchmarks
   3. To optimize hyperparameter settings
   4. To detect whether backdoors were植入 during training

7. Which activity cannot be fully replaced by automated AI detection in a poisoning defense strategy?

   1. Logging model training metrics for audit trails
   2. Human review of suspicious data clusters
   3. Running statistical anomaly detection on data distributions
   4. Generating cryptographic hashes for provenance tracking

8. Why are sophisticated backdoors particularly difficult to detect through standard testing?

   1. They require extremely large datasets to activate
   2. They only affect models with fewer than 10 billion parameters
   3. They trigger on rare, naturalistic combinations developers wouldn't test
   4. They produce audible warnings when activated

9. What does source attestation provide in a data provenance control system?

   1. Provides automatic translation of data between formats
   2. Verifies that the data source certifies its origin and integrity
   3. Guarantees the data was generated by humans
   4. Ensures data is stored in geographic locations meeting compliance requirements

10. Which control area would specifically address the question: 'Do we test the trained model against known trigger patterns?'

    1. Incident response planning
    2. Data provenance tracking
    3. Supply-chain trust
    4. Trigger-pattern evaluation

11. What is the primary purpose of maintaining cryptographic hashes for training data?

    1. To detect any unauthorized modification of data after ingestion
    2. To compress the storage footprint of training datasets
    3. To generate synthetic training examples
    4. To speed up data loading during training

12. In the context of data poisoning, what is a backdoor attack?

    1. A vulnerability in network infrastructure that exposes training data
    2. A physical security breach at data centers
    3. A method for unauthorized access to model weights
    4. An attack where poisoned data causes model behavior changes on specific trigger inputs

13. What should an incident response plan for data poisoning include?

    1. Procedures for model deployment to production
    2. Guidelines for marketing the poisoned model
    3. Methods for increasing training dataset size
    4. Steps to contain and remediate poisoning once detected

14. Why can't provenance tracking alone prevent data poisoning?

    1. Provenance tracking requires more storage than is available
    2. Provenance tracking is incompatible with cloud computing
    3. Provenance only verifies data origin, not whether the source itself was compromised
    4. Provenance cannot be applied to text data

15. Which scenario represents the greatest challenge for automated poisoning detection?

    1. Training data containing obvious duplicate entries
    2. Poisoned data with statistical properties nearly identical to clean data
    3. Training datasets smaller than 1,000 samples
    4. Data sources from well-known major corporations