<p dir="ltr">In recent years, Federated Learning (FL) has emerged as a promising approach for training machine learning models across distributed data sources while preserving privacy. However, traditional FL faces significant challenges in reliabilities, including the risk of the single point of failure and vulnerabilities to adversarial attacks. </p><p dir="ltr">This research proposes an innovative framework, Blockchain-based FL(BCFL), leveraging blockchain to decentralize the FL system and enhance its reliability. To optimize BCFL in resource-constrained environments, we design incentive mechanisms and resource allocation schemes to maximize computational efficiency for clients engaging in both training and mining tasks. Additionally, we introduce a dual-task resource allocation scheme specifically tailored for Mobile Edge Computing (MEC), enabling edge servers to manage both BCFL and offloading tasks efficiently. To address the inherent risk of client dropout in distributed learning, we propose the HieAvg algorithm within a decentralized hierarchical FL framework, mitigating the impact of stragglers through historical weight-based aggregation. This research also introduces the Faker attack, a novel model poisoning approach that exploits weaknesses in similarity metrics commonly used in FL defenses. In response, we develop the Similarity of Partial Parameters (SPP) defense, a random parameter selection strategy that disrupts the predictability of similarity evaluations, offering robust protection against adaptive attacks.</p><p dir="ltr">Our research provides practical strategies to fortify FL systems against reliability vulnerabilities. This work lays the foundation for more secure, reliable, and efficient FL in various environments through decentralized architectures and novel fault </p>