Who Pays for Byzantine Robustness? Effectiveness, Data-Share Dictatorship, and Fairness under Few-Hospital Collaboration
Who Pays for Byzantine Robustness? Validity, Data-Share Dictatorships, and Fairness When Only a Few Hospitals Collaborate
The paper conducts an empirical audit of eight standard Byzantine-robust aggregation rules and finds failures and fairness problems in medical federated learning coalitions of only three to six hospitals with vastly different data scales. The authors run 54,214 training runs on three cross-institutional medical datasets from the FLamby benchmark and use exact twin held-out controls to decompose performance degradation into honest-data loss and attack damage. Results show that when a single hospital holds more than half the training data, sample-size weighting turns the median and GeoMed into a dictatorship of that hospital's clean model; under a dominant-center attack, weighted median, GeoMed and Centered Clipping collapse; under an attack by the largest hospital, the failure of count-based rules stems mainly from honest-data loss (7.5 to 10.4 percentage points on ISIC) rather than adversarial perturbation (2.3 to 3.8 percentage points), and weight-dependent rules suffer true attack damage of 16.1 to 25.1 percentage points on ISIC; unweighted rules persistently discount the institutions whose parameter updates are least typical, penalizing the smallest hospital on Heart and larger institutions on TCGA and ISIC.