We developed a course on differential privacy, using applied exercises to teach the theoretical concepts. The course is designed for graduate students in computer science, focusing on practical applications of differential privacy in data analysis and machine learning.

Course outline:

The course consisted of weekly lectures and in-class assignments, completed using notebook templates we provided. Each topic was accompanied by a homework assignment in which students applied the concepts learned in class. Exercises included implementing differential privacy mechanisms, analyzing privacy attacks, and generating synthetic databases.

Course Topics:

• Attacks on Privacy:
  • K-anonymity, l-diversity, t-closeness
  • Reconstruction attacks (Naive, linear regression, linear programming)
  • Hypothesis testing and Membership inference attacks
  • Classical Auditing
• DP and Basic Mechanisms:
  • Noise addition mechanisms, ReportNoisyMax, AboveThreshold, Exponential Mechanism
• Private Mechine Learning:
  • Memorization, overfitting, and the connection to differential privacy
  • DP, DP-SGD, for logistic regression, and neural networks
  • One run auditing
• Synthetic Database Generation:
  • SmallDB, MWEM, marginal queries