Tangle Ledger for Decentralized Learning @ MPP 2020
Decentralized Learning makes techniques from Federated Learning applicable in scenarios where no central network authority exists. We propose a Tangle Architecture for open Decentralized Learning networks that is robust against basic poisoning attacks from adversarial nodes.
In our paper for MPP 2020, held in conjunction with IPDPS'20, we show that a distributed tangle algorithm is suited to form a consensus on a machine learning model that is trained on node-local data only.
The distributed nature of the algorithm entails an attack surface for potential adversaries trying to manipulate the trained model. We therefore demonstrate how the tangle algorithm can withstand a certain amount of poisoned updates.
You can register for free for IPDPS'20 to get access to our paper and the accompanying slides. There will also be a Zoom session on Friday, where we will present our paper in more detail. Details on how to join will be published on the MPP website.