ByMI: Byzantine Machine Identification with False Discovery Rate Control

Authors: Chengde Qian, Mengyuan Wang, Haojie Ren, Changliang Zou

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments on both synthetic and real data validate the theoretical results and demonstrate the effectiveness of our proposed method on Byzantine machine identification.
Researcher Affiliation Academia 1School of Statistics and Data Sciences, LPMC, KLMDASR and LEBPS, Nankai University, Tianjin, China 2School of Mathematical Sciences, Shanghai Jiao Tong University.
Pseudocode Yes C. Pseudocode of By MI
Open Source Code Yes The code is available at https://github.com/mywang99/By MI.
Open Datasets Yes Datasets. The MNIST (Le Cun et al., 1998), Fashion MNIST (Xiao et al., 2017) and CIFAR10 (Krizhevsky, 2009) datasets are used to verify the performance of our By MI.
Dataset Splits No For MNIST and F-MNIST, all of the samples in the training set are randomly divided into m + 1 = 150 machines (including one master machine) with an equal sample size n = 400. For CIFAR10, we fix m + 1 = 125 and n = 400.
Hardware Specification Yes All the experiments are conducted on an Ubuntu 20.04 LTS server with 64 Intel(R) Xeon(R) Gold 5218 CPUs @ 2.30GHz, 128G RAM and the R platform with version 4.0.2.
Software Dependencies Yes All the experiments are conducted on an Ubuntu 20.04 LTS server with 64 Intel(R) Xeon(R) Gold 5218 CPUs @ 2.30GHz, 128G RAM and the R platform with version 4.0.2.
Experiment Setup Yes The target FDR level is fixed as α = 0.1. Specifically, for the OOD attack we replace the covariates xi s on Byzantine machines by xi = 0.7xi + εp where εp is from Np(νp, σ2Ip) with νp Rp randomly sampled from the standard multivariate normal distribution and σ = 0.2. For the IPM attack, the Byzantine gradients are assigned as a g, where g = 1 |G| P