Experiments and worked examples
Everything here is practice around the six Parts: run it in the browser, predict first, and explain what you saw. Nothing is compulsory within the timetable.
The lecturer demonstrates Lab 01 (Parts 1–3, slide 33) and Lab 06 (Parts 4–6, slide 32). The threshold and evasion explorers run directly in the slides (Parts 1–3 slide 42, Parts 4–6 slide 12). The other browser labs are self-study and are not compulsory within the lecture timetable.
Which experiment supports each Part?
Slide numbers match the Part pages and the PDF handouts.
| Part | In class | Self-study labs |
|---|---|---|
| Part 1 · ML foundations, slides 12–33 | Demo, slide 33: 01 login-alert training | 01 repeat the demo; 13 image pixels (slide 27, representation) |
| Part 2 · detection and evaluation, slides 34–48 | Demo, slide 42: threshold, base rate and analyst capacity explorer | 05 analyst queue (slides 42–43); 03 phishing (slide 44); 11 file features (slide 45); 04 per-entity anomalies (slide 46) |
| Part 3 · limits and new capabilities, slides 49–72 | Quick checks (53–55), prediction (67), control vote (71) | 09 campaign leakage (slides 51–52); 10 drift (slide 56) |
| Part 4 · adversarial ML, slides 6–18 | Demo, slide 12: frozen phishing detector | 08 evasion (slides 10–12; a different two-feature model); 07 poisoning (slides 15–17); 14 traffic windows (how a fixed model reads its input) |
| Part 5 · privacy, supply chain and LLM applications, slides 19–33 | Demo, slide 32: 06 assistant workflow (architecture on slide 24) | 02 approval boundary (slides 24, 31–32) |
| Part 6 · securing AI systems in practice, slides 34–47 | Hospital group exercise (slides 37–38) | 12 incident evidence (slide 45; optional) |
Labs 15–17 are optional modality comparisons and are not used in the lectures. Numerical learners run locally; assistant and incident examples use explicitly authored simulations. Predict, change one condition, inspect the consequence and explain the limits. Lesson numbers inside the labs refer to the sections of the practice notebook.
Each full lab ends with an explanation question and a focused book or paper connection. You have succeeded when you can explain one controlled comparison and its limits. You do not need to try every slider setting.
All seventeen experiments
For Parts 1–3:
- Train a login-alert triage model
- Learn from image pixels
- Build a phishing detector
- Understand the analyst queue
- Train a file detector and inspect its blind spots
- Investigate unusual traffic
- Split the campaigns to expose evaluation leakage
- Update the baseline, or hide an incident?
For Parts 4–6:
- Evade a frozen phishing detector
- Poison a model: clean versus attacked
- Learn from ordered traffic windows
- Walk through an AI assistant
- Test a security boundary
- Follow the evidence: three possible incidents
Optional audio, video and graph training: