Event-Driven Online Vertical Federated Learning
Authors: Ganyu Wang, Boyu Wang, Bin Gu, Charles Ling
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted a comprehensive regret analysis of our proposed framework, specifically examining the DLR under non-convex conditions with event-driven online VFL. Extensive experiments demonstrated that our proposed framework was more stable than the existing online VFL framework under non-stationary data conditions while also significantly reducing communication and computation costs. 5 EXPERIMENT 5.1 EXPERIMENT SETUP 5.2 RESULT ON STATIONARY DATA STREAM 5.3 RESULT ON NON-STATIONARY DATA STREAM |
| Researcher Affiliation | Academia | 1Western University 2Vector Institute 3Jilin University EMAIL EMAIL EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Event-driven online VFL Algorithm 2 Event-driven online VFL with online gradient descent (OGD-Event) Algorithm 3 Event-driven online VFL on Dynamic Local Regret Algorithm 4 Online VFL with Online Gradient Descent (OGD-Full) Algorithm 5 Event-driven online VFL on Static Local Regret |
| Open Source Code | No | The paper does not contain any explicit statement about open-sourcing the code or a link to a repository for the described methodology. |
| Open Datasets | Yes | Dataset We leveraged the Infinite MNIST (i-MNIST) dataset (Loosli et al., 2007) to assess the performance of the proposed methodologies in the context of VFL. ... Experiments on other practical online learning datasets, including the SUSY and HIGGS datasets, are provided in Appendix D.2. SUSY dataset SUSY (Whiteson, 2014b) is a physics dataset from the UCI repository. HIGGS dataset HIGGS (Whiteson, 2014a) is also a physics dataset from UCI repository. |
| Dataset Splits | No | We followed the standard online learning setting, wherein at each round t, client m received only the corresponding feature of a single sample, rather than a batch. Each trial comprised a total of 2,000,000 non-repetitive samples. ... To convert the i-MNIST dataset into a distributed dataset, each image was first flattened into a one-dimensional vector. This vector was then divided into four equal segments to ensure an even distribution of features across the clients. Each client was assigned one of the four feature partitions, while the server was assigned the label. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., library or solver names with version numbers like PyTorch 1.9). |
| Experiment Setup | Yes | The learning rate η was tuned from {1, 0.1, 0.01, 0.001, ...}. The length of the exponential weighted sliding window for the DLR was tuned from {10, 50, 100, 150}. The activation probability p for the Random activation was selected from {0.25, 0.5, 0.75}. The activation threshold Γ was tuned from { -0.2, 0, 0.2, 0.4, 0.6, 0.8}. ... The loss function used by the server was cross-entropy. |