讨论： Differential privacy in one paragraph, and where it does not fit

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

暂无译文，显示原文。 原文

Concrete implementations and the two hazards named in the article. Libraries beyond Google's: OpenDP (a Rust core with Python bindings), Tumult Analytics, PipelineDP (Google and OpenMined, for Beam and Spark), IBM's diffprivlib and PyDP (Python bindings to Google's C++ library); BigQuery offers a `WITH DIFFERENTIAL_PRIVACY OPTIONS(...)` clause in GoogleSQL, which is the 'query interface' case built into a warehouse. The floating-point hazard has a specific origin: Mironov's 2012 paper showed that the textbook Laplace mechanism implemented with floating-point arithmetic leaks through the low-order bits of the noise, which is why libraries use snapping or discrete (integer-valued) noise distributions rather than `random.laplace()`. The large-epsilon hazard has a public reference point: the US Census Bureau applied differential privacy to the 2020 decennial census and published its privacy-loss budget, which is the kind of 'documented product decision' the article asks for. The local-model variants (Apple's device telemetry and Google's RAPPOR in Chrome) add noise on the device before anything is sent, which fits a different case again: the operator never holds the raw values.