Tolerating Adversarial Attacks and Byzantine Faults in Distributed Machine Learning

To tolerate the outliers, robust statistics have been proposed. We summarize them in Table I. Yin et al. [36] proposed Median 111In the paper, we define the uppercase Median as a GAR solution and the lowercase median as the middle value. However, a recent paper proved that the Median aggregation rule is still under an order-optimal error rate [15]. Blanchard et al. [5] proposed Krum for selecting a valid vector update.