softmew⌗

This is a highly optimized software implementation of a memory management unit. Read more about it on the github

Countering UAS AI-Based Object Detection⌗

At work, one of the projects that I’ve been helping out with has been focused on trying to defeat object detection models that might be used in a drone. The conflict in Ukraine has shown the potential of drones in warfare and computer vision and machine planning is getting advanced enough that people might be removed from the loop in the near future.

Countering computer vision models looks a lot different when you need some kind of adversarial artifact that exists in the real world. Most of the current research on defeating computer vision models assumes at least some control over the pixel values in the image. In the real world, you can have some influence over the pixel values by painting or wrapping an object, that does not give absolute control over specific pixel values.

With that limitation, it was interesting to see how the attacks developed and where the research ended up. It turns out that it is much easier to cause false detections under these constraints than it is to try and hide some kind of detection.

You can see the series of blog posts this research produced here (archive).