基于区块链和隐私计算的联邦学习数据安全聚合方法

With the development of 5G and artificial intelligence, the local data generated by massive distributed intelligent terminals is growing exponentially. In order to promote data sharing while protecting user privacy, federated learning technology realizes distributed model training through parameter aggregation. However, the centralized architecture of traditional federated learning has the risk of single point of failure, and faces privacy disclosure threats such as inference attacks. Therefore, this thesis proposes a data security aggregation strategy that integrates blockchain and privacy computing to solve the privacy leakage and data integrity problems in the transmission and aggregation of model parameters in cross-region multi-terminal scenarios. The solution builds a federated learning architecture with regional server clusters as the core, reduces the risk of single point of failure through distributed storage and computing, and ensures transparency, traceability and immutability of data aggregation combined with blockchain smart contracts. Further, a lightweight privacy protection scheme is proposed, which combines elliptic curve homologous encryption and dynamic threshold Shamir secret sharing technology to realize fragment encryption upload and distributed aggregation of gradient parameters. Theoretical analysis shows that this scheme can effectively resist gradient inversion and collusion attacks under semi-honest and active attack models. The experimental results show that the communication and calculation cost are significantly reduced under the premise of controllable model accuracy loss, and the feasibility and high efficiency of the practical application are verified.