Revisiting Differentially Private Algorithms for Decentralized Online Learning

Authors: Xiaoyu Wang, Wenhao Yang, Chang Yao, Mingli Song, Yuanyu Wan

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct simulation experiments with convex functions to verify the performance of our algorithms.
Researcher Affiliation Academia 1School of Software Technology, Zhejiang University, Ningbo, China 2National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China 3State Key Laboratory of Blockchain and Data Security, Zhejiang University, Hangzhou, China. Correspondence to: Yuanyu Wan <EMAIL>.
Pseudocode Yes Algorithm 1 PD-FTGL; Algorithm 2 Private Sum (Smith & Thakurta, 2013); Algorithm 3 PD-OCG; Algorithm 4 CG (Frank & Wolfe, 1956; Jaggi, 2013)
Open Source Code No The paper does not provide any explicit statement or link regarding the public availability of source code for the described methodology.
Open Datasets Yes Moreover, we use two publicly available datasets letter and poker from the LIBSVM repository (Chang & Lin, 2011), and their details are summarized in Table 2.
Dataset Splits Yes For letter, we construct a larger one including 10n copies of the original data, and evenly distribute them among the n local learners, which implies that T = 150000. For poker, we construct a larger one including n copies of the original data, and distribute them in the same way, which implies that T = 1000000.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments.
Software Dependencies No The paper does not provide specific software names with version numbers or specialized package versions used for the implementation of the algorithms or experiments.
Experiment Setup Yes Specifically, let p and q denote the number of features and classes, respectively. The decision set of is defined as K = {X Rq p | X τ}, where X denotes the trace norm of X and τ = 10. [...] By default, we set n = 9 and use the complete graph... Moreover, following previous studies on DP (Abadi et al., 2016; Kairouz et al., 2021), we substitute the original gradient ft,i(Xi(t)) by a clipping gradient defined as gclip(t, i) = ft,i(Xi(t))/ max{1, ft,i(Xi(t))F } where the constant C is tuned from the set {0.01, 0.1, 1, 10} to obtain the best performance... Additionally, for the parameter h of our algorithms, we multiply the theoretical value by a constant tuned from the set {0.0001, 0.0005, 0.001, 0.005, ..., 10, 50} for the best performance. Other parameters of all algorithms are set as what their corresponding theories suggest.