Clean Training Baseline

1. 01

   Concepts & system design

   60 min· read · map the mechanismtimer

   R19 — Hugging Face Datasets lineage & cards

   R20 — Model Cards for Model Reporting paper

   R45 — NIST AI 100-2 training-time poisoning

   R05 — MITRE ATLAS backdoor techniques

   Read the assigned sections; map the mechanism, threat model and trade-offs.

2. 02

   Project implementation

   120 min· build · integrate · debugtimer

   Contrast training/fine-tuning data poisoning with RAG poisoning. Establish a clean baseline on a small local model/classifier task; record dataset lineage, seed, and checkpoint hashes.

   Run a controlled synthetic trigger/backdoor poisoning experiment on your local toy task. Measure clean accuracy vs. trigger activation rate and false positives. *(Never extrapolate toy metrics as universal foundation-model laws).*

   Build, integrate, debug and improve the increment in the active project.

3. 03

   Security lab

   60 min· attack your own systemtimer

   No separate task text for this block on day 27 — the sourced work sits in the block above. Keep the time box.

   Test the abuse case, failure path, authorization bypass or resource limit.

4. 04

   Verification & evidence

   30 min· test · record evidencetimer

   Reproducible training-poisoning lab notebook/script with clean vs. poisoned evaluation metrics.

   Add or run regression tests; record the 9-point evidence record.

5. 05

   Deliberate practice & Socratic review

   30 min· explain aloud · update gapstimer

   Answer the 8 Socratic questions without notes: What? Why? How? What if? Why not? How can I prove it? Where can it break? What have I learned?

   Re-explain the hardest concept aloud, inspect one edge case, update the gap log.

6. total5 h

• R05MITRE ATLAS↗ atlas.mitre.org

  ML attack tactics/techniques: prompt injection, data poisoning, model extraction, and exfiltration.

• R19Hugging Face Datasets Docs↗ huggingface.co

  Dataset loading, revision pinning, dataset cards, and data provenance tracking.

• R20Model Cards Paper↗ arxiv.org

  Model Cards for Model Reporting (Mitchell et al.): intended use, limitations, training data, and disaggregated evals.

• R45NIST AI 100-2 (Adversarial ML)↗ doi.org

  Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (evasion, poisoning, privacy, abuse).

Read only the named sections needed for today's task. Reading sits inside the 60-minute concepts block; it is not extra time.

• docsHugging Face Datasets Documentation · reference docs
  Datasets - Hugging Face

  Covers load_dataset with revision pinning, dataset cards and streaming, which is how you record dataset lineage and a fixed revision for the clean baseline.

  R19checked 2026-11-11
• articlearXiv (Mitchell et al.) · research paper
  [1810.03993] Model Cards for Model Reporting

  Defines intended use, limitations, training data and disaggregated evaluation as reporting obligations, the template for documenting the baseline's data and checkpoint hashes.

  R20checked 2026-11-11
• docsMITRE · knowledge base
  MITRE ATLAS™

  Catalogues poisoning and backdoor techniques by tactic and life-cycle stage, letting you place your synthetic trigger experiment against real-world attack patterns.

  R05checked 2026-11-11
• videoLeevaExplains · 3 min
  Backdoor (Trojan) Attacks on Neural Networks: BadNets Explained

  Explains the BadNets mechanism of a trigger pattern mapped to a target label, the exact design for your controlled trigger-poisoning toy experiment.

  R05checked 2026-11-11
• videoZero to Builder: Cloud, AI & Blockchain · 10 min
  Data Poisoning and Backdoors in LLM Training Explained

  Contrasts weight-level poisoning during training with retrieval-time corpus poisoning, the distinction the day asks you to make explicit before measuring clean accuracy versus trigger rate.

  R05R45checked 2026-11-11

A lab is complete only when it produces evidence. Installing PyRIT, Garak, Presidio or TransformerLens is not completion. Saved on this device only.