Federated learning on the blockchain with SMPC model verification for healthcare systems protected from poisoning attacks

Abstract

Federated learning (FL) has gained traction in several domains, including as smart cities, sectors reliant on the internet of things (IoT), and smart healthcare systems. This is the end product of the growing awareness of the need to protect user data while using neural networks. Customers may work together on training a global structure utilizing FL, all without gaining access to local training materials. Present FL methods, however, may be easily defeated by malicious assaults. It is difficult to identify and stop damaging changes to the model due to its design. Also, there has been a dearth of studies examining the most up-to-date methods for distinguishing FL from false updates without compromising the privacy of the model. This study presents a distributed learning system that uses SMPC model verification to protect healthcare systems against poisoning assaults. The system is built on the blockchain. To begin, we use an encrypted inference method to inspect the ML models of the FL players and eliminate any compromised ones. Verification at a blockchain node takes place once all participants' local models have been securely integrated. To evaluate our proposed system, we utilized medical datasets equivalent to many trials.