Lecture 23: Privacy and security 2022

Q: What types of private data does data science often collect?

Medical records, census data, browser history, location data, social media data.

Q: What can be learned from credit card metadata?

Sensitive patterns about a person’s location, identity, and shopping behavior.

Q: What did the study by De Montjoye et al. (2015) find about credit card data?

Four spatiotemporal points can reidentify 90% of individuals.

Q: How does knowing transaction price affect reidentification risk?

Increases reidentification risk by 22% on average.

Q: Are women more or less reidentifiable than men in credit card metadata?

More reidentifiable.

Q: How can medical data combined with voter data breach privacy?

Linking datasets can reveal identities even if medical data is anonymized.

Q: What is one method to anonymize location data?

Coarsening latitude/longitude into larger regions like zipcodes.

Q: What is randomized response used for?

Protecting individual privacy in sensitive surveys.

Q: How does randomized response work?

Individuals answer truthfully or randomly according to a known probability to protect privacy.

Q: What inequality is linked to high-probability accuracy in randomized response?

Chebyshev’s inequality.

Q: What is differential privacy?

A method that ensures the addition or removal of a single data point doesn’t significantly affect outcomes.

Q: What does adversarial machine learning study?

How to secure ML models against attacks like input perturbations.

Q: What famous example shows adversarial ML vulnerabilities?

A panda image slightly perturbed to be misclassified by a neural network.

Q: What are three types of attacks on ML models?

Inversion, extraction, and data poisoning.

Q: What is inversion attack?

Reconstructing sensitive input data from model outputs.

Q: What is extraction attack?

Stealing model parameters or training data.

Q: What is data poisoning?

Maliciously injecting bad data into training to corrupt a model.

Q: What is federated learning?

Training models across decentralized devices without transferring raw data