Abstract
This paper studies privacy-preserving weighted federated learning within the secret sharing framework, where individual private data is split into random shares which are distributed among a set of pre-defined computing servers. The contribution of this paper mainly comprises the following four-fold: · In the first fold, the relationship between federated learning (FL) and multi-party computation (MPC) as well as that of secure federated learning (SFL) and secure multi-party computation (SMPC) is investigated. We show that FL is a subset of MPC from the m-ary functionality point of view. Furthermore, if the underlying FL instance privately computes the defined m-ary functionality in the simulation-based framework, then the simulation-based FL solution is an instance of SMPC. · In the second fold, a new notion which we call weighted federated learning (wFL) is introduced and formalized. Then an oracle-aided SMPC for computing wFL is presented and analysed by decoupling the security of FL from that of MPC. Our decoupling formulation of wFL benefits FL developers selecting their best security practices from the state-of-the-art security tools. · In the third-fold, a concrete implementation of wFL leveraging the random splitting technique in the framework of the 3-party computation is presented and analysed. The security of our implementation is guaranteed by the security composition theorem within the secret share framework. · In the fourth-fold, a complement to MASCOT is introduced and formalized in the framework of SPDZ, where a novel solution to the Beaver triple generator is constructed from the standard El Gamal encryption. Our solution is formalized as a three-party computation and a generation of the Beaver triple requires roughly 5 invocations of the El Gamal encryptions. We are able to show that the proposed implementation is secure against honest-but-curious adversary assuming that the underlying El Gamal encryption is semantically secure.
Highlights
T homomorphic encryption (HE) concept of federated learning (FL) rst introduced by McMahan et al is a decoupling of model training from the need for direct access to the raw training data [1]
Since the existence of such a solution is not addressed by MASCOT, we provide an interesting research problem below: Question 3: how to construct an eciet yet secure Beaver triple generator which invokes a constant number of the El Gamal encryptions only?
If FL attains the security in the simulation-based framework, the resulting secure federated learning (SFL) is a subset of secure multi-party computation (SMPC)
Summary
T HE concept of federated learning (FL) rst introduced by McMahan et al is a decoupling of model training from the need for direct access to the raw training data [1]. The datasets dened in the FL framework can be categorized as horizontal, vertical and hybrid types. In the horizontal FL, the feature spaces of datasets among dierent organizations (data owners) are same but not overlapped over the sample spaces [5]; in the vertical FL, the sample spaces of datasets among dierent organizations are same but not overlapped over the feature spaces [6], [7]; in the hybrid FL, both feature spaces and sample spaces of dierent organizations are overlapped [8], [9]. We refer to the reader [10][14] and the references therein for more details
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
