FairVFL: A Fair Vertical Federated Learning Framework with Contrastive Adversarial Learning
Authors: Tao Qi, Fangzhao Wu, Chuhan Wu, Lingjuan Lyu, Tong Xu, Hao Liao, Zhongliang Yang, Yongfeng Huang, Xing Xie
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on three real-world datasets validate that our method can effectively improve model fairness with user privacy well-protected. ... We evaluate model performance and fairness on three real-world datasets. |
| Researcher Affiliation | Collaboration | Tao Qi Tsinghua University ... Fangzhao Wu Microsoft Research Asia ... Lingjuan Lyu Sony AI Tokyo, Japan |
| Pseudocode | Yes | The detailed process of a training round in Fair VFL is in Algorithm 1 (Supplementary). |
| Open Source Code | Yes | Codes are available in https://github.com/taoqi98/Fair VFL. |
| Open Datasets | Yes | The first one is ADULT [24], which is a widely used public dataset for fair ML [52, 20, 35, 25]. ... The third one is Celeb A [30], which is a public face attributes dataset. |
| Dataset Splits | Yes | 20,000 randomly selected data samples are used to construct training and validation dataset, and 10,000 randomly selected data samples are used to construct test dataset. ... We randomly select 100,000 news impressions in the first three weeks to construct the training and validation set and select 100,000 news impressions in the last week to construct the test set. |
| Hardware Specification | No | The paper states 'See Supplementary Information' regarding the total amount of compute and type of resources used, but the specific hardware details (e.g., GPU/CPU models) are not provided in the main text. |
| Software Dependencies | No | The paper mentions using 'Adam algorithm [23]' for optimization and 'dropout technique [38]' but does not specify software or library version numbers (e.g., PyTorch version, specific Adam implementation version). |
| Experiment Setup | Yes | The protected representations for gender and age are 32and 64-dimensional, respectively. The weights of contrastive adversarial loss for different sensitive features are set to 0.25. ... We exploit Adam algorithm [23] for model optimization with 1e-4 learning rate. The size of the mini-batch for model training is set to 32. Besides, for both age and gender, we set Ei to 5 for simplification. We also use the dropout technique [38] with a 0.2 drop probability to alleviate model overfitting. |