A new audit of eight standard Byzantine-robust aggregation rules in federated learning has revealed significant issues when applied to real-world clinical collaborations, which typically involve only a few hospitals. The study found that existing validity checks fail to account for breakdowns in skewed consortia, particularly when one hospital holds a majority of the training data. In such scenarios, sample-size weighting can lead to data-share dictatorships and model collapse, with the loss of honest training data being a more significant factor than adversarial attacks. AI
IMPACT Existing federated learning aggregation rules may not be suitable for small-scale clinical collaborations, potentially hindering secure and fair data sharing.
RANK_REASON Academic paper detailing empirical audit of aggregation rules in federated learning. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
- Byzantine robustness
- centered clipping
- FedAvg
- Fed-Heart-Disease
- Fed-ISIC2019
- Fed-TCGA-BRCA
- geometric median
- Hugging Face
- Krum
- Multi-Krum
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →