Loading...
Please wait, while we are loading the content...
Similar Documents
Building Trusted Federated Learning on Blockchain
Content Provider | MDPI |
---|---|
Author | Oktian, Yustus Eko Stanley, Brian Lee, Sang-Gon |
Copyright Year | 2022 |
Description | Federated learning enables multiple users to collaboratively train a global model using the users’ private data on users’ local machines. This way, users are not required to share their training data with other parties, maintaining user privacy; however, the vanilla federated learning proposal is mainly assumed to be run in a trusted environment, while the actual implementation of federated learning is expected to be performed in untrusted domains. This paper aims to use blockchain as a trusted federated learning platform to realize the missing “running on untrusted domain” requirement. First, we investigate vanilla federate learning issues such as client’s low motivation, client dropouts, model poisoning, model stealing, and unauthorized access. From those issues, we design building block solutions such as incentive mechanism, reputation system, peer-reviewed model, commitment hash, and model encryption. We then construct the full-fledged blockchain-based federated learning protocol, including client registration, training, aggregation, and reward distribution. Our evaluations show that the proposed solutions made federated learning more reliable. Moreover, the proposed system can motivate participants to be honest and perform best-effort training to obtain higher rewards while punishing malicious behaviors. Hence, running federated learning in an untrusted environment becomes possible. |
Starting Page | 1407 |
e-ISSN | 20738994 |
DOI | 10.3390/sym14071407 |
Journal | Symmetry |
Issue Number | 7 |
Volume Number | 14 |
Language | English |
Publisher | MDPI |
Publisher Date | 2022-07-08 |
Access Restriction | Open |
Subject Keyword | Symmetry Information and Library Science Federated Learning Artificial Intelligence Blockchain Smart Contract |
Content Type | Text |
Resource Type | Article |