With the popularization of edge computing and IoT devices, large-scale distributed data training puts higher requirements on privacy protection and system robustness. In this paper, we propose a decentralized federated learning framework called SCDFL, which aims to address both privacy leakage and Byzantine attacks. First, the system randomly selects aggregation nodes and detects malicious nodes through a randomized consensus protocol and reputation-based sampling method in each training round to enhance the aggregation robustness. Secondly, the packet-paired onion routing algorithm is introduced to realize multi-layer encrypted transmission of local model parameters to ensure the transmission privacy of data in P2P networks. Finally, the aggregation results are verified by the voting committee mechanism to ensure that all nodes accept the model update unanimously. Experiments prove that on the MNIST dataset, when the proportion of Byzantine nodes is lower than 50%, the global model accuracy of SCDFL can be stably maintained above 91%; even at a malicious proportion of 30%, the training accuracy can be maintained at 86%, which is significantly better than FedAvg and NPBDFL algorithms. This framework has high practical value and scalability, as it protects privacy and defends against Byzantine attacks without imposing an obvious extra burden on system efficiency.