Researchers developed an Adversarial T-shirt, the first successful adversarial wearable for evading detection of people on the move providing invisibility to artificial intelligence.
In the modern world, with the development of imaging technologies, we are all surveilled through CCTV cameras in all public spaces like streets, shopping malls. It means identifying people everywhere by Artificial Intelligence (AI). To prevent this detection, adversarial designs are developed. This concept is defined as “tricking” object detection algorithm by making it see something different from what’s there, or not seeing anything at all. This is a smart way to bypass these types of surveillance.
Adversarial T-shirt for not to be identified
A team of researchers from Northeastern University, Massachusetts Institute of Technology (MIT), MIT-IBM Watson AI Lab, IBM Research, developed a TPS based transformer to model the temporal deformation of adversarial T-shirt caused by pose change of a moving person. Their aim is also to show ensuring the effectiveness of an adversarial T-shirt in the physical world. For this reason, the team proposed a general optimization framework for the design of an adversarial T-shirt in both single-detector and multiple-detector settings.
Steps of the experiment
- Detecting a single person wearing the proposed adversarial T-shirt
- Detecting a single person detection wearing the adversarial T-shirt generated from the baseline method
- Detecting two moving persons wearing two adversarial T-shirts generated from our method and the baseline method, respectively
- Detecting two moving persons wearing
According to experiments in both digital and physical worlds, the proposed adversarial T-shirt achieves 79% and 63% attack success rates respectively. This experiment gives some ideas about how the adversarial perturbations can be implemented with human clothing, accessories, paint on the face, and other wearables.