Project

The goal of doing a project in this course is to explore data privacy by yourself and find something that interests you and make you feel really excited to investigate deeper. During the process, you will learn and practice to do a comprehensive review, plan on your time to finish a project, and write reports.

You are not restricted to follow any format. Latex or Google doc or anything that can convey the knowledge would suffice. There are two categories of a project:

research: a research project aims to introduce new ideas
evaluation: an evaluation project aims to do a comprehensive evaluation and comparison of a topic, e.g., summarize and evaluate all the existing algorithms (code) for a specific problem
application: applying existing techniques to a new domain/dataset, in a non-trivial way

If you aim to do a research project and you put a lot of effort into, say, organizing and cleaning existing methods that no one ever did before (which is also very important), but you only produce a minor new result, that is probably unfortunate in real research. But it is fine in this class, as we can evaluate on both sides. Similarly, if you find new insights during your evaluation, that is also great. Please mark your contribution from two sides clearly.

You are allowed to work individually or form a group of at most two, but keep in mind the expected amount of work will also double if you work in a group. Students in a group will be given the same grade. There is no preference from my side on the research format, topics, or whether you work individually or in a group.

A good project consists of a good proposal (topic evolves as the project goes on, but the more planning you do at this stage, the better), a comprehensive survey of existing work within the topic, and a complete execution. At any stage, you are very welcome to send your plan or draft to the instructor or TA for feedbacks to keep yourself on track.

Some ideas (and you are very welcome to come up with your own idea):

Evaluate/benchmark existing methods for differentially private machine learning
Incorporate public information during differentially private data release
Differential privacy into new machine learning algorithms (I am not an expert of machine learning; I only know image CNN, NLP, GNN; if you know more and want to apply differential privacy to some machine learning paradigm that no one did before, that would be great)
Privacy attacks of new machine learning algorithms
Generating differentially private code automatically
Building a differentially private system (e.g., an SQL engine or a machine learning training system)
Interplay between differential privacy and fairness or poison attacks