An efficient and privacy-preserving federated learning scheme
| dc.contributor.advisor | Lu, Rongxing | |
| dc.contributor.author | Izadi Yekta, Hadiseh | |
| dc.date.accessioned | 2024-01-04T19:25:52Z | |
| dc.date.available | 2024-01-04T19:25:52Z | |
| dc.date.issued | 2022-08 | |
| dc.description.abstract | Federated Learning is increasingly applied to implement many applications that consist of sensitive and confidential data, such as diagnosing medical conditions, implementing smart homes, and smart cities. Federated learning allows a large number of several user groups to collaborate on model training without having to upload data sets to a central server, avoiding the server collecting confidential user data. Nevertheless, sharing model updates with a third party during the training process can still reveal confidential information, which may contain sensitive personal data such as financial or medical records. Many works have been done to preserve privacy in both centralized and distributed training. However, there are still many issues, such as less accuracy, the degraded utility of the learning results, and increasing computation overhead due to cryptography techniques. We propose an efficient and high accurate privacy-preserving neural network federated learning scheme. Especially we implement two privacy-preserving scenarios based on two homomorphic encryption algorithms to compare the scenarios in terms of efficiency and accuracy. In one scenario we employs Paillier Homomorphic Encryption (PHE),and Advanced Encryption Standard (AES) to encrypt the trained models, and on the other scenario we replace the PHE with Symmetric Homomorphic Encryption(SHE).As proof of our scheme’s privacy-preserving properties, we provide an analysis of its security. Finally, we demonstrate the computational efficiency as well as high accuracy of our work through experimental results. | |
| dc.description.copyright | © Hadiseh Izadi Yekta, 2022 | |
| dc.format.extent | xiv, 86 | |
| dc.format.medium | electronic | |
| dc.identifier.oclc | (OCoLC)1425948338 | en |
| dc.identifier.other | Thesis 11154 | en |
| dc.identifier.uri | https://unbscholar.lib.unb.ca/handle/1882/37629 | |
| dc.language.iso | en | |
| dc.publisher | University of New Brunswick | |
| dc.rights | http://purl.org/coar/access_right/c_abf2 | |
| dc.subject.discipline | Computer Science | |
| dc.subject.lcsh | Machine learning. | en |
| dc.subject.lcsh | Data privacy. | en |
| dc.subject.lcsh | Neural networks (Computer science) | en |
| dc.title | An efficient and privacy-preserving federated learning scheme | |
| dc.type | master thesis | |
| oaire.license.condition | other | |
| thesis.degree.discipline | Computer Science | |
| thesis.degree.grantor | University of New Brunswick | |
| thesis.degree.level | masters | |
| thesis.degree.name | M.C.S. |