Decentralized Federated Learning has emerged as an advanced approach to address the limitations of traditional centralized federated learning, such as single points of failure and trust issues. In addition, employing blockchain technology not only facilitates federated learning but also enhances security. However, existing solutions still struggle to operate reliably in fully non-trusted environments. This paper presents a blockchain-enabled decentralized federated learning system that significantly secures the collaborative training process, from the very beginning to the end, by using robust protection mechanisms. We also focus on performance and lightweight design by leveraging off-chain computation and distributed file storage to handle large-model updates. Moreover, our system was implemented and validated in a simulated Ethereum environment, in the comparison with the baseline and state-of-the-art, demonstrating that even with adversarial participants, correct updates still drive model convergence and continuously improve the overall robustness and trustworthiness of federated learning systems.