Refining adversarial attacks on machine-learning phishing webpage detectors through functionality-preserving HTML manipulations

Research has demonstrated that Machine-Learning Phishing Webpage Detectors, which is vulnerable to malicious changes of the input webpage's HTML code. Unfortunately, the newly suggested assaults have not been very effective so far because they target specific HTML components and don't optimise the employment of the manipulations that have been accepted. To get over these limitations, we created a novel set of granular manipulations that can change input phishing webpage HTML code while keeping its aesthetic appeal and malevolent intent (alterations are rendered-preserving and functional by design). After that, we use a query-efficient black-box optimization technique to choose which alterations to apply in order to evade the target detector. In contrast to the weaker attacks created in earlier work, our results demonstrate that our attacks may completely demolish the performance of state-of-the-art ML-PWD with only 30 questions. This allows for a considerably more equitable evaluation of ML-robustness. PWD's