Revisiting EM-based Estimation for Locally Differentially Private Protocols
Yutong Ye
Network and Distributed System Security (NDSS) Symposium 2025 · Day 1 · Privacy & Cryptography 1
This talk, presented by Yutong Ye on behalf of co-authors, delves into critical improvements for **Expectation Maximization (EM)**-based estimation within **Locally Differentially Private (LDP)** protocols. LDP is a stringent privacy framework designed to protect individual data points even before aggregation, making it vital for sensitive data collection across various domains. While EM methods are commonly employed in LDP for tasks like frequency estimation of categorical or numerical data distributions, they suffer from significant practical limitations, notably **overfitting** and **error accumulation** across numerous categories.
AI review
Technically legitimate differential privacy research with a clear problem statement and a plausible, well-scoped solution. Mixture reduction as EM regularization is a sensible idea with real empirical backing, but it's incremental work in a narrow subfield — not a paradigm shift. The k'/k improvement framing is honest about being a high-level bound rather than a tight result.