Federated Learning (FL) is a distributed machine learning technology that enables multiple users to collaborate in training shared models while protecting data privacy. However, FL is susceptible to poisoning attacks, where malicious local model parameters are uploaded to disrupt the performance and correctness of global models. To resist poisoning attacks, this work proposes a new FL model based on a blockchain. Firstly, blockchain validators are iteratively obtained to generate a blockchain consensus committee to validate local model parameters uploaded by trainers. It can effectively exclude malicious nodes with the low accuracy of local model parameters, thus obtaining high-quality local model parameters quickly. Secondly, based on the obtained high-quality local model parameters, an enhanced aggregation rule is proposed. It randomly selects one from the enhanced aggregation functions in each round of FL, making it difficult for malicious nodes to attack the proposed model due to randomization. The experimental results show that the proposed method has higher accuracy than its peers in defending against both data poisoning attacks and model poisoning attacks.