Project rubric

The course project contributes 25% of the course grade, split across four milestones:

For milestone logistics and expectations, see project milestone guide.

1. Topic check-in (5 points)

Criteria Points Description
Relevance 2 The idea clearly fits the course and identifies a concrete privacy question.
Feasibility 2 The scope is realistic for one semester and the available tools.
Planning 1 The memo or talk identifies an initial path, not just a broad topic area.

2. Proposal (5 points)

Criteria Points Description
Problem framing 2 Clear motivation, setting, and threat model.
Method and baselines 2 Specific plan for what will be implemented, compared, or measured.
Scope control 1 Realistic milestones, risks, and fallback plan.

3. Poster / demo (5 points)

Criteria Points Description
Technical content 2 The core question, method, and evidence are understandable and technically sound.
Break and repair 1 The team shows a reproducible challenge, root cause, and repair or a justified clean/insufficient finding.
Results and honesty 1 The team shows actual results, limitations, and residual risk without overclaiming.
Communication and contributions 1 The presentation is clear and team member roles are visible.

4. Final report (10 points)

Criteria Points Description
Technical correctness 3 Methods match the stated threat model; privacy claims, DP claims, and metrics are technically justified.
Evidence, failure, and repair 3 Results include concrete scale and baselines plus a reproducible challenge, root-cause analysis, repair, and regression evidence.
Transfer and limitations 2 The changed setting tests generalization, and the report states what transferred, what did not, and the residual risks.
Writing and artifacts 2 The report is organized, properly cited, and supported by clear figures, tables, configs, tests, and contribution notes.

Additional notes

  1. Novelty is not required
    • This is an undergraduate course project, not a conference submission.
    • Strong reproduction, benchmarking, or application projects can earn top marks.
  2. Claims should match evidence
    • A narrow but carefully tested claim is stronger than a broad claim with thin evidence.
    • Be precise with differential privacy, RAG privacy, side-channel generalization, and LLM-judge results.
  3. Auditability matters
    • Reports should include enough detail to check the result: counts, splits, seeds, prompts, configs, examples, or logs as appropriate.
    • If code or artifacts are shared, they should support the report without exposing private or sensitive data.
  4. Visuals count as communication
    • Figures and tables should be readable and should make the main comparison or failure case easy to see.
    • Visual polish cannot replace weak evidence, but unclear visuals can make strong work hard to evaluate.
  5. Individual vs. group
    • Individuals may submit narrower projects.
    • Teams of 2 should show either broader experiments, stronger comparisons, or a more polished system.
  6. Contribution statements
    • Every project submission should include a short contribution note for each member.
    • If there is a serious contribution imbalance, grades may be adjusted.