TU Wien:Security, Privacy and Explainability in Machine Learning/2023-06-22 Exam-2023S

6 open questions, 90 minutes time:

- Which different methods are there to preserve an individual's privacy when publishing a data set? Explain two of them in detail!

- Explain the concept of DiRo2C and explain the steps that are involved in the method.

- What is model watermarking? Explain the different approaches based on the type of model access!

- Local substitute models vs global substitute models. Comparison and advantages of local over global.

- Image classifer: You have no access to the training data. Explain which attack strategies can be used!

- You run a data base containing different patient diagnoses and population data. Users can query this data base for different statistics. Which risks to privacy are there? Name specific attacks that are potentially possible!