To achieve efficient collaboration between privacy preservation and semi-supervised training in the decentralized federated learning scenario, this paper proposes a decentralized federated semi-supervised learning with zero-knowledge (DFedSem-ZK) algorithm. The algorithm first designs a zero-knowledge feature code that captures local data feature. By integrating Pedersen commitments with Schnorr proofs, the feature code enables secure feature sharing among clients while ensuring the irrecoverability of local data and the verifiability of the exchange process. Furthermore, we construct an efficient decentralized zero-knowledge label propagation method, which leverages the similarity between feature codes to guide pseudo-label generation. This allows for effective dissemination of label information under strict privacy constraints. Computational complexity analysis demonstrates that the computational overhead of the proposed method is significantly lower than that of homomorphic encryption-based schemes. Extensive experiments conducted on multiple datasets show that the algorithm consistently outperforms existing baselines across varying data distributions and unlabeled data configurations, with notable gains in both accuracy and robustness. Additionally, experimental evaluations on the variable number of clustering cores and network topologies further demonstrate the influence of clustering core selection on model performance, as well as the stability and practicality of the proposed algorithm under different decentralized settings.