With the development of Internet technology, network traffic classification has become a key tool for security risk prevention. Currently, machine learning (ML) and deep learning (DL) based techniques have excellent performance in feature extraction and pattern recognition, but they still face two major challenges: firstly, the problem of ‘Isolated Data Island’ makes it difficult to share data, and centralised training is prone to privacy leakage; secondly, the model relies on a large amount of labelled data, and the high cost of annotation restricts its practical application. In this paper, we propose an asynchronous federated contrastive semi-supervised learning method (FCAL) for network traffic classification, which adopts a federated learning architecture to protect data privacy, and builds an adversarial autoencoder (AAE) and contrastive learning (CL) module to solve the problem of insufficient labelled data. Through adversarial training and contrast learning, FCAL learns fine-grained traffic features across multiple clients and improves classification accuracy through globally shared models. In addition, an asynchronous update mechanism is used to allow clients to flexibly update model parameters according to individualised training situations, reducing communication delay and improving computational efficiency. Experiments show that FCAL outperforms other baseline models on two types of public datasets with fewer labelled samples, and the classification performance and computational efficiency are significantly improved, verifying its effectiveness in distributed network traffic classification.