A Method for Stealing Traffic Detection Models Based on Equivalent Feature Sets

Machine learning models have shown excellent performance in the field of traffic detection. However, they also face new security challenges. This study focuses on model stealing attacks in the traffic detection field. Existing model stealing methods are difficult to implement in real - world network environments. They either cannot be applied to the traffic detection field or require prior knowledge of the target model's feature set. To achieve the stealing of traffic detection models in a truly black - box scenario and overcome the strong assumption of relying on the known target model feature set in existing methods, this study proposes a method for stealing traffic detection models based on feature inference. This study introduces the concept of “equivalent feature set”, analyzes the prediction logic of the target model and the characteristics of traffic data through two feature inference algorithms, constructs an equivalent feature set, and uses it as the feature set for training a substitute model. The results of stealing experiments on multiple target models show that in most experimental settings, the stealing rate exceeds 85%, with a maximum of 96.64%, demonstrating the high efficiency and stability of the method. At the same time, experimental verification shows that the equivalent feature set can accurately capture key features, significantly improving the stealing effect of the substitute model, with an improvement amplitude of more than 40.48%.