Robust Aggregation
A core technique in federated learning for defending against Byzantine faults. It replaces arithmetic averaging with methods such as coordinate-wise median, trimmed mean, and Krum to identify and suppress anomalous data uploaded by malicious nodes.
Timeline (last 90 days)
Classical robust aggregation rules such as coordinate-wise median, trimmed mean, and Krum suppress the influence of extreme values through statistically robust estimation.
Robust aggregation methods may misjudge legitimate updates from normal clients as anomalous under non-independent and identically distributed (non-IID) data, because differences in data distributions among clients themselves cause gradient directions to diverge.
The algorithm replaces naive averaging with robust aggregation at the central server to identify and suppress anomalous uploads that deviate from the normal distribution.
Common robust aggregation methods include coordinate-wise median, trimmed mean, Krum, and its variants.
Theoretical guarantees for robust aggregation typically rely on the assumption that the proportion of malicious nodes is below a certain threshold (e.g., no more than half of the total nodes).
All Facts (5)
Classical robust aggregation rules such as coordinate-wise median, trimmed mean, and Krum suppress the influence of extreme values through statistically robust estimation.
50%UnverifiedRobust aggregation methods may misjudge legitimate updates from normal clients as anomalous under non-independent and identically distributed (non-IID) data, because differences in data distributions among clients themselves cause gradient directions to diverge.
50%UnverifiedThe algorithm replaces naive averaging with robust aggregation at the central server to identify and suppress anomalous uploads that deviate from the normal distribution.
50%UnverifiedCommon robust aggregation methods include coordinate-wise median, trimmed mean, Krum, and its variants.
50%UnverifiedTheoretical guarantees for robust aggregation typically rely on the assumption that the proportion of malicious nodes is below a certain threshold (e.g., no more than half of the total nodes).
50%