Test-Time Robust Personalization for Federated Learning

Authors: Liangze Jiang, Tao Lin

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate the advancement of Fed THE+ (and its degraded computationally efficient variant Fed THE) over strong competitors, for training various neural architectures (CNN, Res Net, and Transformer) on CIFAR10 and Image Net and evaluating on diverse test distributions. Along with this, we build a benchmark for assessing the performance and robustness of personalized FL methods during deployment.
Researcher Affiliation Academia Liangze Jiang2, , Tao Lin1, liangze.jiang@epfl.ch; lintao@westlake.edu.cn 1Research Center for Industries of the Future, Westlake University 2EPFL
Pseudocode Yes Pseudocodes are deferred to Appendix B.
Open Source Code Yes Code: https://github.com/LINs-lab/Fed THE.
Open Datasets Yes Datasets. We consider CIFAR10 (Krizhevsky et al., 2009) and Image Net (Deng et al., 2009) (downsampled to the resolution of 32 (Chrabaszcz et al., 2017) in our case due to computational infeasibility)
Dataset Splits Yes We then uniformly partition each client s data into local training/validation/test sets.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running the experiments.
Software Dependencies No The paper mentions optimizers (SGD, SGDm, Adam) and Group Norm but does not specify software library names with version numbers (e.g., PyTorch 1.9, TensorFlow 2.x, Python 3.x).
Experiment Setup Yes For all the experiments, we train CIFAR10 with a batch size of 32 and Image Net-32 with that of 128. For head ensemble phase (personalization phase) of Fed THE and Fed THE+, we optimize the head ensemble weight e by using a Adam optimizer with initial learning rate 0.1 (when training CNN or Res Net20-GN) or 0.01 (when training CCT4), and 20 optimization steps.