The anonymity of the darknet makes it a frequent platform for illegal activities, making the characterization and classification of darknet traffic crucial to maintaining network security. Past research has commonly employed machine learning models and deep learning models for darknet traffic classification and characterization, but these models have exhibited low accuracy and F1 scores. Furthermore, many darknet traffic classification methods do not inherently consider the adversarial nature of their deployment environment, making it equally important to defend against adversarial attacks targeting darknet traffic classifiers. Therefore, this paper proposes a machine learning-based darknet traffic classification method that incorporates new feature values during the data pre-processing stage, which significantly contribute to the performance of the machine learning classifier. To enhance the robustness of the machine learning classifier, this paper proposes a two-layer defense mechanism consisting of a detector and a denoiser. The detector leverages Natural Scene Statistics (NSS) to characterize adversarial attack traffic, enabling the detection of such attacks. The denoiser, implemented using an autoencoder, projects the adversarial attack traffic detected by the detector back to its benign manifold. The proposed method is evaluated on the CIC-Darknet-2020 dataset, and its defense mechanism is assessed using four adversarial attacks. The experimental results show that the darknet traffic classification method proposed in this paper outperforms its comparison models, achieving an accuracy of 98.52% and an F1 score of 98.53%. Furthermore, the proposed defense method can identify adversarial attacks with an accuracy of over 99%, effectively improving the robustness of darknet traffic classifiers.