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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Mitigating the Privacy–Utility Trade-off in Decentralized Federated Learning via f-Differential Privacy
Authors: Xiang Li, Chendi Wang, Buxin Su, Qi Long, Weijie Su
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on synthetic and real datasets demonstrate that our methods yield consistently tighter (ϵ, δ) bounds and improved utility compared to Rényi DP based approaches, illustrating the benefits of f-DP in decentralized privacy accounting. |
| Researcher Affiliation | Academia | Xiang Li University of Pennsylvania EMAIL Buxin Su University of Pennsylvania EMAIL Chendi Wang Xiamen University University of Pennsylvania EMAIL Qi Long University of Pennsylvania EMAIL Weijie Su University of Pennsylvania EMAIL |
| Pseudocode | Yes | Algorithm 1: Decentralized DP-SGD Algorithm 2: DECOR: DP-SGD with Corr. Noise |
| Open Source Code | Yes | The codes are available at https: //github.com/lx10077/PN-f-DP. |
| Open Datasets | Yes | The first is a logistic regression model trained on a binarized version of the UCI Housing dataset,3 and the second is an MNIST image classification [38] task using a simple convolutional neural network. |
| Dataset Splits | Yes | The training data is then distributed across n = 28 users, each holding 64 local samples, according to an expander graph topology. We compare three approaches: decentralized DP-GD, local DP-GD, and our random-walk-based DP-GD under different noise variances. We focus on the privacy loss from user 1 to user 2, fixing (ϵ, δ) = (10, 10 5), and record both objective values and test accuracy every 100 iterations. |
| Hardware Specification | No | All experiments are conducted on a CPU cluster with 200 GB of memory. |
| Software Dependencies | No | The Py Torch implementation of the model is shown below. |
| Experiment Setup | Yes | For fair comparison, we fix the total iterations T, local updates K = 1, and noise variance σ2 = 1. We set T = Θ( log n λ2 ), which is the number of steps required for the random walk to converge to a minimal precision level. We use the numerical composition method in [29, 36] to convert PN-f-DP to (ϵ, δ)-DP. |