VBSFL: A Robust Blockchained Split-Fed Learning Model for Secured Distributed Learning

George Aziz, Sajib Mistry, Aneesh Krishna

Abstract: Split-fed learning (SFL) is a novel approach within distributed collaborative machine learning that combines federated learning (FL) and split learning (SL). While SFL benefits from FL's speed and SL's efficiency, it also inherits their disadvantages, including trust issues such as model poisoning and training-hijacking attacks, and additional reliability concerns such as a single point of failure and lack of motivation. Existing solutions have integrated blockchain technology to address all reliability issues and poisoning attacks, but are limited to FL. Moreover, there are limited solutions to address training-hijacking, such as the SplitGuard protocol. In this article, we propose two solutions, validated blockchained split-fed learning (VBSFL) and VBSL, focusing on VBSFL, which leverages blockchain technology by building on VBFL and incorporating the SplitGuard protocol to address these challenges. Experimental results with real-world datasets demonstrate the effectiveness, efficiency, and scalability of the proposed approach.