Computer Vision News - December 2018

put it into a model, and the model tells whether the picture is a cat, right? And that’s not really useful for this problem because most of the imagery that is relevant for identifying brands is more two-dimensional logo graphics kind of imagery. There is a fair bit of work to just apply what’s obviously very well treated in the literature on standard computer vision to this kind of use case. ” Their technology requires a very specific domain application of a general set of techniques that have been perfected in the academic literature over the last 10 years. Transferring that sort of technique to this domain has been quite challenging, and computer vision, although a powerful tool, does not solve this problem sufficiently on its own. In general, it needs to corroborate with a whole host of other signals extracted from the mail. The software uses a lot of domain-specific email techniques, many of which do not involve machine learning such as approximate string matching or heuristics. Developing their technology requires experts in both computer vision and email. Because not many people have this intersecting kind of work experience, it has taken a long time for computer vision to bleed over into mail protection . “ Imagine taking just the standard, off the shelf TensorFlow computer vision model like any of your practitioners would do. Now you give it an email. Now what? That whole process of applying the off the shelf stuff to the mail involves a couple of interesting problems ” shares Baggett. A fraudulent email may or may not even contain any imagery at all. Baggett says that they often come across an email containing what looks like a logo but is actually just HTML. For instance, an email that impersonates the retailer, Target, will have a big white word, Target, on a red background, which looks quite convincing. That means Inky must deal with the process of rendering HTML. Pauline Luc 19 Computer Vision News Inky - Phishing Detection Application