Lossless Compression of Efficient Private Local Randomizers
Authors: Vitaly Feldman, Kunal Talwar
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Figure 1. Expected ℓ2 2 error of mechanisms Priv HS, Priv Unit, Priv Unit Optimized and SQKR for values of ε between 1 and 8. These plots show both Priv Unit and Priv Unit Optimized are more accurate than Priv HS and SQKR in the whole range of parameters |
| Researcher Affiliation | Industry | Vitaly Feldman 1 Kunal Talwar 1 1Apple. |
| Pseudocode | Yes | Algorithm 1 R[G, γ]: PRG compression of R ... Algorithm 2 PI-RAPPOR randomizer ... Algorithm 3 Server-side frequency for PI-RAPPOR |
| Open Source Code | No | The provided link is for an implementation of a Kashin-based mean estimation scheme, which is discussed as a related work and a baseline for comparison, not the authors' own methodology described in the paper. |
| Open Datasets | No | The paper describes empirical comparisons and uses parameters like d, n, and ε for these comparisons, but does not provide concrete access information (link, DOI, etc.) to a specific dataset used for training or evaluation in their experiments. |
| Dataset Splits | No | The paper discusses statistical settings and error metrics but does not specify any training, validation, or test dataset splits. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used to run the experiments. |
| Software Dependencies | No | The paper mentions an implementation from (Kas) for comparison but does not list specific software dependencies with version numbers for their own described methodology or experiments. |
| Experiment Setup | Yes | We show error bars for the empirical squared error based on 20 trials. ... For d = 1,000, n = 10,000 and ε taking integer values from 1 to 8. ... The Priv Unit algorithm internally splits its privacy budget ε into two parts ε0, ε1 = 1 ε0. ... optimize the splitting so as to minimize the variance proxy, by evaluating the expression for the variance proxy as a function of the θ = ε0/ε, for 101 values of θ = 0.00, 0.01, 0.02, . . . , 0.99, 1.0. |