Project (Spring 2026)
All project-related requirements are organized under this project section. Presentation milestones (pitch, milestone report, and poster) are part of the same project workflow.
Project scope
The goal of this project is to explore data privacy in large language models (LLMs) and/or agentic systems. Every project must align with this theme.
You may choose one of these formats:
- Research: Introduce a new idea or method.
- Evaluation: Build a careful benchmark and comparison study.
- Application: Apply known methods to a new domain or realistic setting.
These categories overlap in practice. Strong projects usually combine literature review, implementation, and critical analysis.
Team policy
- You may work individually or in a group of up to 2.
- Group projects should have roughly double the total work compared with an individual project.
- Group members submit the same materials and receive the same grade.
Milestones at a glance
| Milestone |
Week(s) |
Format |
Weight |
| Project pitch |
6-7 |
Presentation (7 min incl. Q&A) |
10% |
| Proposal document |
9 |
Written plan (1-2 pages) |
10% |
| Milestone report |
12-13 |
Presentation (15 min) |
10% |
| Poster session |
16 |
Poster presentation |
15% |
| Final report |
16 |
Written report |
15% |
For presentation details, see presentation guide.
For detailed grading criteria, see rubric.
Suggested project directions
- Training data leakage and memorization in LLMs.
- Privacy-preserving fine-tuning or alignment (e.g., DP-SGD vs PEFT).
- RAG privacy (retrieval logs, embeddings, and leakage controls).
- Agentic systems privacy (tool-use logs, action traces, policy constraints).
- Private inference for LLMs (HE/MPC/TEE latency-cost tradeoffs).
- Auditing deployed LLM/agent systems via risk benchmarks or red teaming.
- LLM fine-tuning: HuggingFace Transformers + PEFT/TRL (optionally Opacus).
- RAG / agents: LlamaIndex or LangChain + vector store (e.g., FAISS).
- Serving / inference: vLLM or llama.cpp for reproducible inference benchmarks.
Proposal and final report template
Use this structure for your proposal, then expand it for the final report.
- Title + Team: project title, members, roles.
- Motivation + Use Case: problem statement and impact.
- System/Model Setting: model stack, data, deployment context.
- Threat Model + Privacy Goals: adversary assumptions and privacy target.
- Baselines: systems/methods used for comparison.
- Your Approach: what you build, change, or evaluate.
- Evaluation Plan: datasets, metrics, and experiments.
- Milestones + Timeline: planned execution through course checkpoints.
- Risks + Mitigations: expected blockers and backup plans.
Deliverable expectations (high level)
- Proposal: clear scope, feasibility, timeline, and evaluation plan.
- Poster: concise visual communication of motivation, method, and results.
- Final report: complete technical write-up with analysis, limitations, and citations.
For point-by-point grading, use Project rubric.