Privacy-Preserving Federated Convex Optimization: Balancing Partial-Participation and Efficiency via Noise Cancellation

Authors: Roie Reshef, Kfir Yehuda Levy

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

Reproducibility Variable Result LLM Response
Research Type Experimental We ran Algorithm 1 on MNIST using a logistic regression. The parameters are G = 2 785 = 39.6, L = 785/2 = 392.5, D = 0.1, which brings us S = 118.1. Our model has d = 10 785 = 7850 parameters. We compared our algorithm (called Our Work ) to SGD with noise, inspired by (Abadi et al., 2016) (called Noisy SGD ), and to the other work (Lowy & Razaviyayn, 2023) (called Other Work ). We kept the same parameter in all 3 algorithms to the best of our abilities, and in all tests the total data samples used across all machines is n = 60,000. ... We compare both the test accuracy and running time. For our first experiment, we fix m = 50, M = 100, and compare various values of ρ. We show our results in Table 1.
Researcher Affiliation Academia 1Faculty of Electrical and Computer Engineering, Technion, Haifa, Israel. Correspondence to: Roie Reshef <EMAIL>, Kfir Yehuda Levy <EMAIL>.
Pseudocode Yes Algorithm 1 DP-µ2-FL with Partial-Participation
Open Source Code No The paper does not contain any explicit statement about releasing source code or a link to a code repository.
Open Datasets Yes We ran Algorithm 1 on MNIST using a logistic regression.
Dataset Splits No The paper mentions using MNIST and a total of 60,000 samples but does not specify how the dataset was split into training, validation, or test sets.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions 'logistic regression' and 'SGD' but does not specify any software libraries or their version numbers (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes We ran Algorithm 1 on MNIST using a logistic regression. The parameters are G = 2 785 = 39.6, L = 785/2 = 392.5, D = 0.1, which brings us S = 118.1. Our model has d = 10 785 = 7850 parameters. ... For our first experiment, we fix m = 50, M = 100, and compare various values of ρ.