What's fascinating here is AdFlush is a classical feature engineering approach: define a bunch of features on the data manually, and then use ML to figure out the most useful / impactful ones. This is not the "throw terabytes of data and see what happens" approach we see with LLMs. It's a bit funny to even point this out because I don't recall the last time a feature-engineered ML project made it to the HN front page.
Features can be brittle, but they are understandable. The paper's appendix [1] lists the 27 features that will likely make a request/resource "ad-related". These include interesting ones like JS AST depth, average JS identifier length, the "bracket to dot notations ration in JS", and a number of graph measures for the graph of scripts.
And contrary to what comments in this thread are saying, they do compare against a blocklist-based adblocker: uBlock Origin. That's in section 5.5. They say they outperform uBlock Origin. But even they say they don't reduce overall page time bc their algorithm is expensive.