An Empirical Study of Adversarial Samples on the Performance of Unknown Attack Detection Models

With the continuous development of computer technology, the methods of network attacks are evolving. The emergence of an increasing number of novel network attack types has led to network attack detection models based on a variety of technologies. Among them, machine learning or deep learning attack detection models have achieved superior performance in real engineering environments compared to traditional intrusion detection systems. However, in the field of cybersecurity, existing network attack detection models based on machine learning or deep learning are highly susceptible to adversarial attacks. Attackers can significantly influence the output of trained machine learning models by introducing minor perturbations to the original input data, thereby reducing the detection performance of these models. Based on theoretical analysis, this study investigates and implements four network attack detection models constructed using different open-set recognition techniques. The performance of these models in detecting unknown attacks is tested on plaintext datasets, encrypted datasets, and darknet dataset. Furthermore, this paper explores and implements three gradient-based adversarial attack generation methods. These methods are applied to different models on various datasets to conduct adversarial attacks, and the changes in model performance before and after the attacks are evaluated, which is useful to identify and understand the vulnerabilities of existing network attack detection models and contributes to the research on improving model robustness and interpretability.