FNBench
收藏资源简介:
FNBench是一个基准研究,旨在评估联邦学习(FL)在面对噪声标签时的鲁棒性。该数据集考虑了三种不同的噪声标签模式,包括合成标签噪声、不完美的人为标注错误和系统性错误。评估涉及了五种图像识别数据集和一个文本分类数据集,并纳入了十八种最先进的方法。数据集的创建旨在解决联邦学习中数据质量保证的问题,特别是当客户端的本地数据集可能包含不同程度的标签噪声时。FNBench为联邦学习社区提供了一个实验平台,以便研究人员可以测试和比较不同方法的性能。
FNBench is a benchmark study designed to evaluate the robustness of Federated Learning (FL) against noisy labels. The test datasets adopted by this benchmark cover three distinct noisy label patterns, including synthetic label noise, imperfect human annotation errors, and systematic errors. The evaluation involves five image recognition datasets and one text classification dataset, and incorporates eighteen state-of-the-art methods. The development of FNBench aims to address the challenge of data quality assurance in federated learning, particularly when local datasets held by clients may contain varying degrees of label noise. FNBench provides an experimental platform for the federated learning community, enabling researchers to test and compare the performance of different methods.

- 1FNBench: Benchmarking Robust Federated Learning against Noisy Labels中国科学院计算技术研究所 · 2025年



