fedlib.aggregators.Clippedclustering

class Clippedclustering(agg='mean', max_tau=100000.0, linkage='average')

Bases: object

A robust aggregator from paper “An Experimental Study of Byzantine-Robust Aggregation Schemes in Federated Learning”. It separates the client population into two groups based on the cosine similarities.

Parameters:

tau (float) – threshold of clipping norm. If it is not given, updates are clipped according the median of historical norm.