Decentralized Learning using Hashgraph Consensus

Federated learning has become increasingly popular for its ability to process large, distributed datasets and speed up learning while protecting data privacy. However, it typically relies on a central server for coordination, which can be a bottleneck and a single point of failure. To address these limitations, we developed a novel distributed learning architecture that eliminates the need for a central server. The architecture utilizes the hashgraph consensus algorithm (HCA), a distributed ledger technology, which enables the computing nodes to train machine learning models using local data and aggregation with models received from their neighbors. Our work demonstrates that distributed learning using hashgraph consensus can be performed efficiently and is a valid alternative to traditional federated learning. To strengthen this claim, we analyze resilient federated learning in a decentralized setting. Our analysis includes scenarios with denial-of-service and model poisoning attacks. We introduce trimmed soft-medoid (TSM), a resilient aggregation method that has proven resilience to model poisoning attacks. It can be performed at every node using the information available from the hashgraph. An extensive evaluation is conducted using two multimodal machine learning tasks, human emotion recognition and activity recognition. The results confirm that decentralized learning using hashgraph consensus maintains performance parity with traditional federated learning using a central server. This is shown in both normal and adversarial scenarios. We also evaluate the latency and memory overhead of the proposed approach. These are reported to be under an acceptable range, latency of 1s and memory overhead of 8.8-13 GB, for decentralized machine learning.