Research on Network Performance Trade-offs in the Integration of Blockchain and Federated Learning: Architecture, Mechanisms, and Optimization

With the surge in the number of Internet of Things (IoT) devices and the increasingly stringent data privacy regulations, Federated Learning (FL), as a privacy-preserving distributed machine learning paradigm, has gained widespread attention from academia and industry. However, traditional federated learning relies on a central server for model aggregation, which has inherent defects such as single point of failure (SPoF), lack of transparency, and vulnerability to security threats such as poisoning attacks[1] and Sybil attacks[2] , where malicious nodes create multiple fake identities to compromise the global model. Blockchain technology, with its decentralized, immutable, and traceable characteristics, has been introduced to solve the aforementioned trust and security issues, forming blockchain-assisted federated learning (BC-FL). Although BC-FL enhances the security and robustness of the system, the introduction of blockchain inevitably brings significant overhead to network performance. There is an essential contradiction between the consensus mechanism delay, block propagation bandwidth consumption, and complex cryptographic operations of blockchain and the requirements of federated learning for high-frequency communication, low latency, and fast convergence[3].This paper aims to provide a detailed review and in-depth theoretical analysis of the network performance trade-offs in the BC-FL system. First, we construct the hierarchical system architecture of BC-FL and propose an improved scheme based on directed acyclic graph (DAG) and sharding. Second, based on the research of Zhang et al.[3], we establish an end-to-end mathematical model of latency and communication overhead and quantify the impact of different consensus mechanisms on the training cycle. Third, we deeply analyze the performance requirements differences in two typical scenarios: Internet of Vehicles (IoV) and Internet of Medical Devices (IoMT) and demonstrate the experimental results of customized architectures (such as CPRE-IoV and Med-Q Ledger). The experimental results show that the optimized BC-FL architecture can reduce the model update latency to less than 200ms while ensuring security and maintaining convergence accuracy comparable to traditional FL.