Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Numerical Accounting in the Shuffle Model of Differential Privacy
Authors: Antti Koskela, Mikko A. Heikkilä, Antti Honkela
TMLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper we show how numerical accounting (Koskela et al., 2020; 2021; Gopi et al., 2021) can be employed for privacy analysis of both single and composite shuffle DP mechanisms. We demonstrate that thus obtained bounds can be up to orders of magnitudes tighter than the existing bounds from the literature. We also evaluate how significantly adversaries with varying capabilities differ in terms of the resulting privacy bounds using the k-randomised response mechanism. Figure 1 shows a comparison between the PLD and RDP applied to the pair of distributions P and Q given in Equation 3.1. As we see from Figure 3, this approach leads to considerably lower ε(δ)-bounds than the approach by Girgis et al. (2021). |
| Researcher Affiliation | Collaboration | Antti Koskela EMAIL Nokia Bell Labs University of Helsinki Mikko Heikkilä mikko.a.heikkila@helsinki.fi Department of Computer Science University of Helsinki Antti Honkela antti.honkela@helsinki.fi Department of Computer Science University of Helsinki |
| Pseudocode | No | The paper describes methods and refers to Algorithm 1 of Doroshenko et al. (2022) but does not present a new pseudocode block within its text. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing code or a link to a code repository. |
| Open Datasets | No | The paper focuses on theoretical models and numerical privacy accounting. It does not use or provide access information for any specific public dataset for empirical validation. |
| Dataset Splits | No | The paper conducts numerical analysis and comparisons of privacy bounds rather than experiments on a dataset that would require training/test/validation splits. |
| Hardware Specification | No | The paper mentions Monte Carlo integration for calculations but does not specify the hardware on which these computations were performed. |
| Software Dependencies | No | The paper refers to methods like 'FFT-based method' and 'Algorithm 1 of Doroshenko et al. (2022)', but it does not specify any software names with version numbers (e.g., Python, PyTorch, CUDA, or specific solvers). |
| Experiment Setup | Yes | Figure 1: Evaluation of δ(ε) for general single and composite shuffle ε0-LDP mechanisms using RDP accounting and FFT-based numerical accounting (PLD) applied to the pair of distributions P and Q given by the post-processing result of Feldman et al. (2023). Number of users n = 104 and the LDP parameter ε0 = 4.0. Figure 2: ... we set ε0 = 3.0, n = 104, nc = 2000, subsampling ratio q = 0.01, α1 = exp( −0.25), αnα = exp(0.25), and take a logarithmically equidistant α-grid. Figure 4: k-RR with the strong adversary As ... Here n = 1000, probability of randomising γ = 0.25, and k = 4. |