<p>Federated learning, characterized by its reliance on a central server, is vulnerable to single points of failure and security concerns stemming from the inherent trust placed in the server. Blockchain technology offers a robust solution to these challenges. Nevertheless, federated learning systems built on blockchain are not immune to threats like poisoning attacks and membership inference attacks. To tackle these challenges, this paper proposes a novel secure framework for blockchain-based federated learning. Based on the framework a client-side local model update uploading mechanism and a global model aggregation strategy are developed. The local model update uploading mechanism combines client grouping with secure multi-party computation protocols, enabling training nodes to aggregate group-level model updates and upload them to the blockchain without leaking any individual update information. Additionally, the grouped updates help to obscure each client’s model updates, effectively defending against inference attacks. The global model aggregation strategy utilizes committee nodes to score the grouped updates, and smart contracts select high-quality updates based on these scores to form the global model. This approach can separate malicious model updates from benign ones, mitigating the impact of poisoning attacks and ensuring the quality of the global model. In this framework, the blockchain serves not only as a medium for decentralized communication, but also as a trust-free coordination and governance infrastructure that maintains participant reputation, constrains the contribution process, and prevents malicious manipulation–thereby enabling the system to operate stably in weak-trust or even adversarial environments. Experimental results demonstrate that the proposed framework can effectively defend against attacks from malicious nodes while ensuring the security and reliability of the model.</p>