Adversarial Attacks on Featureless Deep Learning Malicious URLs Detection

Detecting malicious Uniform Resource Locators (URLs) is crucially important to prevent attackers from committing cybercrimes. Recent researches have investigated the role of machine learning (ML) models detect URLs. By using ML algorithms, first, features URLs are extracted, and then different trained. The limitation this approach that it requires manual feature engineering does not consider sequential patterns in URL. Therefore, deep (DL) used solve these issues since they able perform featureless detection. Furthermore, DL give better accuracy generalization newly designed URLs; however, results our study show models, such as any other can be susceptible adversarial attacks. In paper, we examine robustness demonstrate importance considering susceptibility before applying detection systems real-world solutions. We propose a black-box attack based on scoring functions with greedy search for minimum number perturbations leading misclassification. examined against types convolutional neural networks (CNN)-based URL classifiers causes tangible decrease more than 56% reduction best classifier (among selected work). Moreover, training shows promising reducing influence model less 7% average.