Automatically detecting vulnerable websites before they turn malicious

Significant recent research advances have made it possible to design systems that can automatically determine with high accuracy the maliciousness of a target website. While highly useful, such systems are reactive by nature. In this paper, we take a complementary approach, and attempt to design, implement, and evaluate a novel classification system which predicts, whether a given, not yet compromised website will become malicious in the future. We adapt several techniques from data mining and machine learning which are particularly well-suited for this problem. A key aspect of our system is that the set of features it relies on is automatically extracted from the data it acquires; this allows us to be able to detect new attack trends relatively quickly. We evaluate our implementation on a corpus of 444,519 websites, containing a total of 4,916,203 webpages, and show that we manage to achieve good detection accuracy over a one-year horizon; that is, we generally manage to correctly predict that currently benign websites will become compromised within a year.