A Robust Decentralized Federated Aggregation for Heterogeneous Data via Blockchain System

Blockchain-based federated learning (FL) technology enhances the performance of traditional FL and is widely applied in various fields, such as image recognition and medical diagnosis. To improve system security, numerous studies typically focus on threats like Byzantine attacks and privacy leakage. However, real-world data is often heterogeneous, and existing blockchain-based FL systems still suffer from: 1) Byzantine detection algorithms often perform poorly in heterogeneous data scenarios and 2) traditional noise addition methods to protect privacy may lead to reduced data availability. To address the above challenges, we propose SraFBS, a Byzantine-robust federated aggregation framework based on a blockchain system for heterogeneous data in decentralized scenario. Our framework introduces a role allocation mechanism, categorizing users to workers, a verification committee, and an aggregation committee. Workers use a key negotiation mechanism to create masks that can be offset during the aggregation phase, protecting privacy while reducing the impact of noise. The verification committee identifies byzantine attackers, and the aggregation committee ensures secure aggregation. Experimental results on image classification tasks show that our system effectively identifies 40% Byzantine attackers in heterogeneous data scenarios, achieving a successful attack rate below 4%, while maintaining model training accuracy close to attack-free FedAvg.