Group Fair Federated Learning via Stochastic Kernel Regularization
Authors: Huzaifa Arif, Pin-Yu Chen, Keerthiram Murugesan, Alex Gittens
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through experiments on standard benchmark datasets across both IID and Non-IID settings for regression and classification tasks, KFFL demonstrates its ability to balance accuracy and fairness effectively, outperforming existing methods by comprehensively exploring the accuracy fairness trade-offs. |
| Researcher Affiliation | Collaboration | Huzaifa Arif EMAIL Rensselaer Polytechnic Institute, Troy, NY, United States Keerthiram Murugesan EMAIL IBM Research, Yorktown Heights, NY, United States Pin-Yu Chen EMAIL IBM Research, Yorktown Heights, NY, United States Alex Gittens EMAIL Rensselaer Polytechnic Institute, Troy, NY, United States |
| Pseudocode | Yes | Algorithm 1 Federated Proximal Gradient Descent (Fed Prox Grad) ... Algorithm 2 KFFL Client Side ... Algorithm 3 KFFL Server Side |
| Open Source Code | Yes | Code is available at github.com/Huzaifa-Arif/KFFL. |
| Open Datasets | Yes | We conduct extensive experiments on standard benchmark datasets under both IID and Non-IID data distributions on both classification and regression tasks... For classification tasks, we used five datasets commonly encountered in group fairness research (Han et al., 2023): Adult, COMPAS, Bank, ACS, and German... When the underlying task is regression, we incorporate additional datasets into our evaluation. Beyond the Adult dataset, we also consider the Law School and Communities and Crime datasets, as utilized in the work Agarwal et al. (2019a). |
| Dataset Splits | Yes | We evaluate the performance in two different federated learning settings: IID (independent and identically distributed) and Non-IID. In the IID setting, each client is provided with an equal number of samples and a shared local data distribution Li et al. (2020). In the Non-IID setting, each client has a different distribution of the protected attribute. Specifically, since the protected group A is binary with attributes being A0 and A1, half of the clients have 90% of A0 and 10% of A1, while the other half has 90% of A1 and 10% of A0 Li et al. (2020). |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions the use of 'Adam Optimizer' and the 'Pyrfm library' but does not provide specific version numbers for these software components or any other key software dependencies. |
| Experiment Setup | Yes | The batch size is uniformly set to 64. Each algorithm undergoes a total of 10 global training rounds, with each round comprising 5 local epochs on every client. The experiments are conducted with 4 clients... The learning rate α was set to 0.01 and Adam Optimizer Kingma & Ba (2014) was used for optimization. The dimensionality of the feature maps used for kernel approximation, denoted as D, is set to 10. The fairness parameter λ controls the trade-off between optimizing predictive performance and enforcing fairness... Each point represents a different fairness weight λ ranging from 20.00 to 1000.00 for both KFFL and KFFL-TD using the ADULT dataset and from 0.01 to 123.16 for the COMPAS dataset. |