One of the series of experiments with AI-generated fooling patches.
Curious about tech?
Our company is engaged in the analysis and recognition of static images and video stream for more than 7 years. Our applications are used for detecting people in videos, using products in advertising, etc.
Through recent years adversarial attacks on machine learning models have become an increasing interest. By making only minor changes to the input of a convolutional neural network, the output of the network can be changed to produce a completely different result. The first attack was conducted by slightly changing the pixel values of an input image to trick classifier, to infer a wrong class.
Other approaches tried to find out “patches” that can be applied to objects to fool detectors and classifiers. Some of these approaches also showed these attacks are possible in reality, for instance by modifying an object and shooting it with a video camera. However, all these approaches are focused on classes that practically do not contain an interclass variety (for example, brake lights). In this case, the object’s known structure is used to create an adversarial patch on top of it.
We present an approach to creating adversarial patches for purposes with a large amount of intraclass diversity – people. The goal is to create a patch that can successfully hide a person from a detector.
We found out that our system can significantly reduce the accuracy of human detection. Our approach also works well enough in real-world scenarios, where patches are filmed by cameras. As far as we know, we are the first who tried to attack targets with such a high level of intraclass diversity – people. The image generation system that we use to create T-shirt patches is built on our own designs and inspired by the ideas of the Cornell University group (https://arxiv.org/abs/1904.08653).
In real tests with printed patches, we have found out that they work well enough while hiding people from surveillance cameras. We assume security systems that use such detectors will also be useless in detecting people wearing lothing with these patches applied.
We will gratefully accept any suggestions and comments. Please share with us your experience of using our clothes in real life. If you have new ideas on our project please write us an email: tech[at]hidexa.com.