Invisibility cloak

Overview

This paper studies the art and science of creating adversarial attacks on object detectors. Most work on real-world adversarial attacks has focused on classifiers, which assign a holistic label to an entire image, rather than detectors which localize objects within an image. Detectors work by considering thousands of “priors” (potential bounding boxes) within the image with different locations, sizes, and aspect ratios. To fool an object detector, an adversarial example must fool every prior in the image, which is much more difficult than fooling the single output of a classifier.

In this work, we present a systematic study of adversarial attacks on state-of-the-art object detection frameworks. Using standard detection datasets, we train patterns that suppress the objectness scores produced by a range of commonly used detectors, and ensembles of detectors. Our ultimate goal is to build a wearable “invisibility” cloak that renders the wearer imperceptible to detectors.

Making an Invisibility Cloak: Real World Adversarial Attacks on Object Detectors

You can purchase your own invisibility clothes below. Proceeds will benefit the College Park Foundation and other charities selected by the authors. Note: sizes tend to run small.

Invisibility Cloak Store @ RageOn

  

This stylish pullover is a great way to stay warm this winter, whether in the office or on-the-go. It features a stay-dry microfleece lining, a modern fit, and adversarial patterns the evade most common object detectors. In this demonstration, the YOLOv2 detector is evaded using a pattern trained on the COCO dataset with a carefully constructed objective.

Video Demo

While a full-scale demo had been delayed due to COVID, here’s a short composite of some of our test footage.

Approach

We load images from the COCO detection dataset, and pass them through a detector. When a person is detected, and pattern is rendered over that person with random perspective, brightness, and contrast deformations. A gradient descent algorithm is then used to find the pattern that minimizes the “objectness scores” (confidence in the presence of an object) for every object prior.

  

image

image

image

Thanks

Thanks to Facebook AI for their support on this project!