01 / EXTRACT
Model Extraction
Train surrogate models on the first 50, 200 and 800 outputs from the same fixed sequence; compare agreement with task accuracy.
CS442 · Interactive learning
CS442
Course and Lab
Explore trustworthy machine learning through human–computer interaction: form testable hypotheses, operate real models, inspect feedback and explain results. Turn experiment output into supported judgments through teacher, student and AI assistant collaboration.
ICA4 · About 70 minutesModel Extraction and Membership InferenceCoding Workbench · AI Teaching Assistant
In-class activity 04
ICA4 · Model audit and AI interaction
Groups save an original prediction, then code model extraction and membership inference in the browser. The workbench shows actual query and training status; formal evaluation returns group metrics and signed receipts. Each group advances to the next stage through its own submissions.
02 / INFER
Membership Inference
Explore thresholds on public shadow data and observe TPR/FPR. Query baseline and regularized models and audit them using both scores.
03 / INTERPRET
Interpretation and discussion
Explain your judgment and evidence to the AI assistant, check explanations after teacher questions, record revisions and submit a conclusion with limits.
70-minute classroom schedule
| Time | Activity | Expected output |
|---|---|---|
| 0—5 minutes | Sign in, check access scope and budget, and save an original prediction | Threat model and experiment hypothesis |
| 5—30 minutes | Code three extraction budgets and formally evaluate | Budget comparison and extraction receipts |
| 30—50 minutes | Explore shadow thresholds and complete four membership audits | Threshold observations and membership inference receipts |
| 50—60 minutes | Ask follow-up questions and explain differences with teacher guidance | Result analysis and revised notes |
| 60—70 minutes | Complete analysis, submit a conclusion and download evidence | Group conclusion and verifiable results |
Teacher, student and assistant discussion
Students propose and test
State predictions and reasons first, operate the experiment interface, then record measured metrics and observed differences.
Model feedback and assistant hints
The experiment model returns measured results; the AI assistant uses evidence available to your group to offer explanations, hints and follow-up questions.
Teachers question and give feedback
Review group progress, notes and conversations. Help students distinguish evidence from inference and state where their conclusions apply.
By default, each group can send 50 questions per classroom, with at most 600 output tokens per answer. Teachers can adjust the question limit. Remaining quota appears on the student interaction page.
Participation and submission
- Sign in with the classroom token from your teacher. Use the Coding Workbench directly for the main operations.
- Record hypotheses, measured metrics, explanations and limits in Lab Notes, preserving how your judgment changes.
- After formal evaluation, compare three extraction budgets and four membership audits; support your conclusion with server receipts.
- Complete the analysis and enter submission. Write a conclusion of at least 120 characters including numerical evidence, a model comparison and a limitation.
- Download and save your group evidence JSON. Groups using the Notebook route should also save their Notebook and worksheet. Submit as your teacher requires.