As the proportion of encrypted traffic increases, it becomes increasingly challenging for network attacks to be discovered. Although existing methods combine unencrypted statistical features, e.g., average packet length, with machine learning algorithms to achieve encrypted malicious traffic detection, it is difficult to escape the influence of artificially forged noise, e.g., adding dummy packets. In this study, we propose a novel encrypted malicious traffic detection method named RobustDetector (RD) for obfuscated malicious traffic detection. The core of the proposed method is to use the dropout mechanism to simulate the process of original features being disturbed. By introducing noise during the training phase, the robustness of the model is improved. To validate the effectiveness of RobustDetector, we conducted extensive experiments using public datasets. Our results demonstrate that RobustDetector achieves an average F1-score of 90.63% even when random noise is introduced to the original traffic with a probability of 50%. This performance underscores the potential of our proposed method in addressing the challenges of obfuscated malicious traffic detection.