Fine-Tuning Personalization in Federated Learning to Mitigate Adversarial Clients
Authors: Youssef Allouah, Abdellah El Mrini, Rachid Guerraoui, Nirupam Gupta, Rafael Pinot
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We support our findings with empirical results on mean estimation and binary classification problems, considering synthetic and benchmark image classification datasets. (Abstract) and 2.2 Experimental Validation and 3.3 Experimental Validation (Section Headers) |
| Researcher Affiliation | Academia | Youssef Allouah EPFL Lausanne, Switzerland Abdellah El Mrini EPFL Lausanne, Switzerland Rachid Guerraoui EPFL Lausanne, Switzerland Nirupam Gupta University of Copenhagen Copenhagen, Denmark Rafael Pinot Sorbonne Université and Université Paris Cité, CNRS, F-75005 Paris, France |
| Pseudocode | Yes | Algorithm 1 Interpolated Personalized Gradient Descent for client i C |
| Open Source Code | Yes | We give open access to our code as well as clear instructions on how to reproduce our experimental results. (NeurIPS Paper Checklist, Question 5) |
| Open Datasets | Yes | We empirically investigate the impact of Byzantine adversaries on the generalization performance, in our personalization framework, using the Phishing dataset (Chiew et al., 2019) and the MNIST dataset (Le Cun and Cortes, 2010). |
| Dataset Splits | No | The paper states 'We subsequently sample the test datasets using the same class distribution as the train datasets for each client, and we evaluate the trained models on these local datasets.' However, it does not provide explicit details about training/validation/test splits with percentages or counts, or mention a separate validation split. |
| Hardware Specification | Yes | We run our experiments on a server with the following specifications: HPe DL380 Gen10 2 x Intel(R) Xeon(R) Platinum 8358P CPU @ 2.60GHz 128 GB of RAM 740GB ssd disk 2 Nvidia A10 GPU cards (Appendix D.1 Compute) |
| Software Dependencies | No | The paper describes the model architecture (e.g., 'Convolutional Layer (1, 32, 5, 1) + Re LU + Maxpooling'), but does not explicitly list software dependencies with specific version numbers (e.g., 'PyTorch 1.9'). |
| Experiment Setup | Yes | We use T = 100 and the learning rate η = 0.05. (Appendix D.2 MNIST) and We fixed the following default values: n = 600, f = 100, m = 20, σ = 15, σh = 2 (Section 2.2) |