Evasion Attacks and Defense Mechanisms for Machine Learning-Based Web Phishing Classifiers

Abstract

Phishing is still a big problem on the internet. People who want to do things make fake websites that look a lot like real ones. They do this to trick people into giving away information like passwords. Even though we have gotten better at stopping these websites with things, like lists of bad sites and lists of good sites and even using computers to help us phishing is still getting worse. The people who make these sites are always coming up with new ways to avoid getting caught by the systems that are supposed to stop them. Phishing attacks are changing all the time. Recent studies have revealed that machine learning-based phishing classifiers are particularly vulnerable to evasion attacks, where adversaries manipulate website features to mislead prediction models while preserving the malicious functionality and visual appearance of phishing pages. This study looks at ways to trick machine learning-based web phishing classifiers. Finds a good way to defend against these attacks to make the system more robust. The framework we came up with looks at the parts of URLs and webpage content and uses many machine learning algorithms to figure out if something is phishing. We make samples by changing the important parts that the classifier looks at but we make sure the website still works and looks the same. We check if the pages we manipulate still look okay by using things like Mean Squared Error to see how different they are, from the thing. Furthermore, a resemblance-based detection approach is introduced to identify evasion-induced phishing attacks. The experimental results demonstrate that evasion attacks significantly degrade classifier performance, highlighting the need for robust defense strategies. The proposed defense mechanism effectively detects adversarial phishing samples and improves classification resilience. This research contributes to strengthening machine learning-based phishing detection systems against adversarial threats, ensuring enhanced cybersecurity in modern web environments.