FeDis: Federated Learning Framework Supported by Distributed Ledger.

Despite the several advantages that distributed ledgers provide to end-users and the system, it also gives access to all data in the chain to every user with a participating ledger node, even if encrypted. While modifying data in the blockchain is a difficult and complex process, the sole existence of data in this way raises some security concerns. To address this issue, this article proposes FeDis, a framework that combines federated learning with a distributed ledger. In the proposed FeDis framework, federated learning is used to train a global model by aggregating local gradient models trained by the participants in the learning process. One of the key features of the framework is the possibility of handling heterogeneous data and devices from different participants. By combining federated learning with a distributed ledger, local model weights are saved in the ledger after encryption and then fetched by the server to be aggregated and to create a global model. The use of a distributed ledger in FeDis increases user trust by providing an additional layer of security and traceability. In addition, the framework is tested using the NASA Bearing dataset showing that the extra layer of security does not compromise the accuracy of the local models.