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..
Accelerated Vertical Federated Adversarial Learning through Decoupling Layer-Wise Dependencies
Authors: Tianxing Man, Yu Bai, Ganyu Wang, Jinjie Fang, Haoran Fang, Bin Gu, Yi Chang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that Dec VFAL ensures competitive robustness while significantly achieving about 3 10 speed up. Experimental evaluation on multiple datasets demonstrates that Dec VFAL achieves 3 10 speedup compared to existing VFAL methods while maintaining competitive robust accuracy, making robust VFL training practically feasible for real-world deployment. |
| Researcher Affiliation | Academia | 1School of Artificial Intelligence, Jilin University, China 2Western University, Canada 3International Center of Future Science, Jilin University, China 4Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, MOE, China EMAIL {gwang382}@uwo.ca EMAIL |
| Pseudocode | Yes | Algorithm 1: Dec VFAL algorithm |
| Open Source Code | Yes | Detailed procedures are available in Appendix C, and the code is accessible at https://github.com/workelaina/Dec VFAL.git. Our code is available at the anonymous repository https://anonymous.4open.science/r/Dec VFAL-F5E4/ as mentioned in Section 6. |
| Open Datasets | Yes | We utilized three public datasets instead for our main experiments: MNIST [34], CIFAR-10 [31], large scale dataset CIFAR-100 [31], and Tiny-Image Net [33] (Appendix C.6). |
| Dataset Splits | Yes | MNIST [34]: A benchmark dataset for image classification, comprising 60,000 examples for training and 10,000 examples for testing. Tiny-Image Net [33]: ... contains 200 classes with 500 training images, 50 validation images, and 50 test images per class. |
| Hardware Specification | Yes | The hardware specifications are detailed in Table 13. Table 13: Hardware Specifications Experiment Description CPU GPU MNIST Robust Training AMD EPYC 7551P A4000*1 CIFAR-10 Robust Training AMD EPYC 7452 4090*4 CIFAR-100 Robust Training AMD EPYC 7452 4090*4 Performance across various NN architectures Intel E5-2683 v4 4090*1 Impact of split position AMD EPYC 7J13 4090*4 Impact of the number of modules AMD EPYC 7J13 4090*4 Impact of the number of the clients Intel Platinum 8336C 4090*8 Limitation of the setting of M and N Intel Fold 6430 4090*8 |
| Software Dependencies | Yes | In our experiments, we utilized the following software environment: Py Torch version 2.2.1, CUDA version 12.1, and Python version 3.11. |
| Experiment Setup | Yes | All models were trained to convergence using Adam optimizer with a fixed learning rate of 0.0001 for fair comparison. Detailed parameter settings and hardware specifications for the training procedures are summarized in Appendix C.3 and Table 13. Table 11: Hyperparameters for Adv. Training |