Certifiably Byzantine-Robust Federated Conformal Prediction

Authors: Mintong Kang, Zhen Lin, Jimeng Sun, Cao Xiao, Bo Li

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we demonstrate the robustness of Rob-FCP against various portions of malicious clients under multiple Byzantine attacks on five standard benchmark and real-world healthcare datasets. We empirically evaluate Rob-FCP against multiple Byzantine attacks.
Researcher Affiliation Collaboration 1University of Illinois at Urbana-Champaign, USA 2Carle s Illinois College of Medicine, USA 3GE Healthcare, USA 4University of Chicago, USA.
Pseudocode Yes An overview of Rob FCP is presented in Figure 2, with the pseudocode detailed in Algorithm 1 in Appendix F. Algorithm 1 Malicious client identification
Open Source Code Yes The codes to reproduce all the evaluation results are publicly available at https://github.com/ kangmintong/Rob-FCP.
Open Datasets Yes We evaluate Rob-FCP on a variety of standard datasets, including MNIST (Deng, 2012), CIFAR-10 (Krizhevsky et al.), and Tiny-Image Net (Le & Yang, 2015). Our evaluation of Rob-FCP also cover two realistic healthcare datasets: the Sleep Heart Health Study (SHHS) dataset (Zhang et al., 2018) and a pathology dataset Path MNIST (Yang et al., 2023).
Dataset Splits Yes In this work, we focus on the split conformal prediction setting (Papadopoulos et al., 2002), where the data samples are randomly partitioned into two disjoint sets: a training set Itr and a calibration (hold-out) set Ical = [n]\Itr. For SHHS: 2,514 patients (2,545,869 samples) were used for training the DNN, and 2,514 patients (2,543,550 samples) were used for calibration and testing.
Hardware Specification Yes The valuation is done on a RTX A6000 GPU.
Software Dependencies No The paper does not provide specific software dependencies with version numbers for reproducibility.
Experiment Setup Yes Without specification, the desired coverage level 1 \alpha is set 0.9. By default, we set \beta to 0.5 to establish a consistent level of data heterogeneity.