The many faces of AI in the Phishing-website landscape

Seminar University of St. Gallen


Saint Gallen, Switzerland

Oneliner: What are some ways in which AI can be used in the context of phishing websites?

Phishing websites are everywhere. This fact may come at a surprise when considering the thousands of papers proposing artificial-intelligence (AI) techniques to counter this threat. Some of these techniques “work”, i.e., they can reliably detect phishing websites—which is clearly an encouraging result. However, many “state-of-the-art” AI methods can also be trivially fooled with little effort by naive attackers—which is clearly a disheartening result. Finally, AI methods can also be offensively used by attackers to circumvent AI-based detectors—which is clearly a worrying result.

In this talk, I will explore these three complementary classes of results, each denoting a different “face” of AI. Specifically, I will explain on how AI can be used to catch phish. Then, I will show how to trivially evade these AI-based methods with simple modifications that anyone could do. Finally, I will reveal more sophisticated—but still affordable—ways to maliciously use AI tools to circumvent phishing detectors powered by AI. During this journey I will also emphasize the role of the end-user: ultimately, a phishing website must deceive a human—not an AI.

(Shoutout to Katerina Mitrokotsa, who “hosted” me for this talk! It was great to see her again after meeting her for the first time in Dagstuhl.)

Slides Venue

((Another version of this talk has been given to a seminar at UniMoRe on December 17th, 2024))