Hasso-Plattner-Institut
Prof. Dr.-Ing. Bert Arnrich
 

Fully-Decentralised Machine Learning

Bjarne Pfitzner

Joint work with Jossekin Beilharz and Robert Schmid from the Operating Systems and Middleware group.

Motivation

Federated learning is an effective and potentially privacy-preserving approach to train machine learning models on private, distributed (health) datasets. Since federated learning relies on a central entity to facilitate the training process and aggregate the global model, there is the possibility that this central server is compromised or malicious, risking vulnerability to different types of attacks.

To prevent this attack surface altogether, we propose resorting to fully-decentralised learning. Here, the data owners communcate directly within a peer-to-peer network without having to trust an outsider. Previous works on fully-decentralised machine learning have used single-chain distributed ledger techologies, such as a blockchain to store the continuously updated model weights. The consensus mechanism of the chosen ledger is then used to find the optimal model over time.

We propose to extend this setup to a directed acyclic graph (DAG) instead of a chain, where newly added models (or nodes in graph terminology) refer back to their parent models on the DAG.

Decentralised Learning with Random Tip Selection

Our fully-decentralised learning approach uses the tangle architecture known for its use for the IOTA cryptocurrency [1]. Data owners can improve the global model at any time by loading the most current version of the tangle and performing two random walks through the DAG, starting with the very first so-called genesis node.

The random walks go in the opposite direction of the edges and are performed until two so-called tips are found, which are nodes without an incoming edge. Now that the client has found two very recent models, he/she can take a simple average of them and consecutively update the model for a number of epochs. In the next step, the newly found model is compared to the current reference model on the tangle, which is the current best model chosen by global consensus. If the new model performs better than the reference model (on the clients' local data), it is allowed to be added to the tangle, referring back to the two models that were averaged in the beginning. This way, the tangle grows over time and the reference model improves for every participant. We could show that this fully-decentralised approach achieves comparable performance to regular federated learning. 

Implicit Specialisation

As an extension to the fully-decentralised learning approach on a DAG, we adjusted the random tip selection procedure to no longer be random. Instead, clients evaluate potential next models during the walk thorugh the DAG, always taking the route of the better-performing model. The rest of the approach stays the same. This way, if clients possess very different data, they implicitly build clusters on the DAG that correspond to specialised models for their data cluster. This approach is more beneficial than splitting the clients before training, because firstly, it may not be possible to split the clients due to data privacy issues, and secondly, in the beginning of the training process, the clusters do not build yet, because the everyone can benefit from each others updates. Thus, the models for each cluster can perform better than models that were trained using only the cluster-specific clients in the first place.

Resilience against Poisoning Attacks

As another benefit, the so-called accuracy tip selection procedure also improves the resilience against model poisoning attacks. They describe an attack where an adversary poses as a benign participant of a learning system, but instead of submitting useful models, they submit completely random ones, of even ones that are trained for a specific misclassification. Due to the accuracy tip selection, these malicious models are very rarely selected and included into the DAG, so that they get stale quickly and do not count towards the network consensus.

References

  1. Serguei Popov, The tangle, 2018. [Online]. Available: https://iota.org/IOTA_Whitepaper.pdf.

Publications

  • Implicit Model Specialization through Dag-Based Decentralized Federated Learning. Beilharz, Jossekin; Pfitzner, Bjarne; Schmid, Robert; Geppert, Paul; Arnrich, Bert; Polze, Andreas in Middleware ’21 (2021). 310–322.
     
  • Tangle Ledger for Decentralized Learning. Schmid, R.; Pfitzner, B.; Beilharz, J.; Arnrich, B.; Polze, A. (2020). 852–859.