Noisy SIGNSGD Is More Differentially Private Than You (Might) Think

Authors: Richeng Jin, Huaiyu Dai

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, extensive experiments are conducted to validate the theoretical results. In this section, we present experimental results to validate our theoretical analyses in the previous sections.
Researcher Affiliation Academia 1College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, China. 2Department of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC, USA.
Pseudocode Yes Algorithm 1 G-Noisy Sign Compressor (Chen et al., 2020b) and Algorithm 2 Differentially Private Noisy SIGNSGD
Open Source Code No The paper does not contain any explicit statement about open-sourcing the code for the methodology described, nor does it provide a link to a code repository.
Open Datasets Yes Datasets and Models: We evaluate the performance of the algorithms on two commonly used benchmarks for differentially private distributed learning: Fashion-MNIST (Xiao et al., 2017) and CIFAR-10 (Krizhevsky et al., 2009).
Dataset Splits Yes The Fashion-MNIST dataset consists of 60,000 training samples and 10,000 testing samples. Each sample is a 28 × 28 size gray-level image. The CIFAR-10 dataset contains 50,000 training samples and 10,000 testing samples. Each sample is a 32 × 32 color image.
Hardware Specification No The paper does not explicitly describe the hardware used to run its experiments, such as specific GPU or CPU models. It only mentions general experimental settings without hardware details.
Software Dependencies Yes Our experiments are mainly implemented using Python 3.8 with packages Numpy 1.19.2 and Pytorch 1.10.1.
Experiment Setup Yes Hyperparameters: The per-example gradient clipping thresholds are set to C = 1 and C = 2 for Fashion-MNIST and CIFAR-10, respectively. We tune the learning rate from the set {0.001, 0.002, 0.003, 0.005, 0.01, 0.02, 0.03, 0.05, 0.1, 0.2, 0.3, 0.5, 1, 2, 3, 5, 10} and run the algorithms for 500 communication rounds. We run all the algorithms for 10 repeats and present the mean test accuracy. We train both neural networks from scratch with a batch size of 32 in our experiments.