Privacy for Distributed Computing
Stub. Achieving privacy when no single trusted party collects all the data: Warner's randomized response (the original local-privacy mechanism, predating differential privacy, for truthfully surveying a sensitive attribute) and Kairouz et al.'s treatment of local differential privacy and private federated/distributed data collection.
Corresponds to CS/ECE 598DA, Week 6. The mechanisms in differential-privacy-mechanisms-and-lower-bounds assume a trusted curator holds the raw data and only releases noised outputs (the central model of DP). This note covers the alternative: each individual randomizes their own data before anyone else ever sees it (the local model), going back to Warner’s randomized response (a survey technique where a respondent flips a coin to decide whether to answer truthfully or give a fixed response, giving plausible deniability to any individual answer while still allowing the aggregate truth to be estimated). Kairouz et al. covers the modern generalization: local differential privacy and privacy-preserving protocols for federated/ distributed data collection, where no party ever holds the full dataset in the clear.
Outline (TODO — flesh out each)
- Randomized response (Warner)
- Local differential privacy
- Private federated and distributed data collection (Kairouz et al.)