Don't need to go that far, i think. If you had your extension hash some piece of each keyframe (basically: tokenize some IDs for each keyframe) and submit them to a database you could then see which parts were shown to everyone vs only to some people and only display those. Basically similar to how sponsorblock crowd sources its sponsor segment detection but automated. Some people would see the ads but then you'd know what the og video was unless it gets edited.

This is assuming they're not reencoding the video for each advertisement, which they probably aren't. If they are it probably gets easier, actually. Sponsorblock could do that.