> So it looks like they might have already been using PhotoDNA. Of course they were working with NCMEC (and therefore using PhotoDNA, which is owned by NCMEC);

> So it looks like they might have already been using PhotoDNA. Of course they were working with NCMEC (and therefore using PhotoDNA, which is owned by NCMEC); it would have been legal suicide for them not to. They were doing this for years, before they were even acquired by Yahoo. > Maybe I'm misunderstanding, but PhotoDNA seems to be a tool for identifying reposts and edits of already known illegal images -- those that have previously been found and had their hashes entered into the database. So if Tumblr had a problem with original content (which could include underage users posting explicit pictures of themselves, as well as the worse things we think of when we hear the phrase 'child porn'), I don't think that would help. How many people do you think have Tumblr accounts where they post only new, never-before-seen pornography of underage children, which has never been posted on any other blog before, and never once post a single photo that has been previously identified as child pornography[0]? Of those, how many do you think are able to ensure that none of their followers ever repost/reblog that photo on any other blog which also happens to contain at least one other photo that's been previously identified as child pornography? Of those, how many do you think are able to ensure that none of their followers have been previously identified as highly-connected nodes in the underground networks dedicated to sharing child pornography, and therefore pretty much exclusively post child pornography or follow people who they believe are likely to post child pornography? Of those, how many do you think are able to ensure that nobody who ever sees one of those photos ever decides to download it and upload it as an attachement to an unsent email in their Gmail drafts folder, or put it in a private Dropbox folder (shared with nobody), or sends it on any of the many "cloud" services which also actively monitor for child pornography and do the same sorts of graph analysis to identify people who are using their services to store or share child pornography? Again, once you understand how these underground networks work, and once you realize that this is mostly a problem of social graph analysis and not image recognition/classification, you realize that it's very easy to solve. [0] Remember, not just new to Tumblr, but new to everyone who is working with NCMEC/ICMEC (which means every large and not-so-large company in the entire world that hosts user-provided content).