In the field of network security, Trojan traffic detection faces challenges brought by the complexity of traffic data and the scarcity of annotations, which limits traditional supervised learning methods. To this end, this paper proposes a semi-supervised learning method that combines virtual adversarial training (VAT) and Mean-Teacher model to effectively utilize unlabeled data and improve the performance of Trojan traffic detection. This method enhances the robustness of the model to uncertain inputs through VAT, does not need to rely on real tags, is suitable for semi-supervised learning scenarios, and significantly improves the ability to distinguish between normal traffic and Trojan traffic. At the same time, the Mean-Teacher model optimizes the learning efficiency of unlabeled data through the collaboration of teacher and student models. The teacher model uses an exponential moving average (EMA) to update parameters, and the student model imitates teacher predictions. Experimental results show that the model demonstrates excellent detection performance under different label ratios, and can still maintain high detection accuracy especially when labels are scarce. By combining VAT and Mean-Teacher model, this paper provides an efficient solution for Trojan traffic detection, breaking through the limitations of traditional supervised learning and demonstrating its potential in practical network security applications.