Two men were walking on a street side-by-side towards a camera. One of them was wearing a black outfit and another, white T-shirt with Khakis. The astonishing part was that the camera reports only one person, the other person with white T-shirt was invisible.
The white T-shirt was designed by researchers at Northeastern University, IBM, and the Massachusetts Institute of Technology, to deceive the specific neural network analysing the video. An assistant professor of electrical and computer engineering at Northeastern, Shelly Lin said, “Deep neural networks are very powerful, but also can be vulnerable to adversarial attacks. When you wear a T-shirt, it is highly possible that the deep neural network won’t identify you in the image.”
These deep neural networks mentioned are artificial intelligence that identifies, analyse, visualise, and classify the sounds, images, or other inputs. The researchers usually train these algorithms by giving exposure to the environment until they can identify voices or images (like a selfie).
However, the main obstacle with deep neural networks was that how they will identify between a human or horse or an umbrella. Shelly Lin and his colleagues were analysing and studying these deep neural networks and working to train for object detectors so that they can categorise and enable him to pick out correctly and label them in a video with “person”, “horse” or “umbrella”.
However, the researchers wanted to make an image that is undetectable. In a digital era, it is very common to manipulate the specific pixels and confuse the neural networks but in the real world, it was harder. Still, researchers left no stone unturned.
They observed that on a stop sign or when the body was stationary, an artificial intelligence system was working and could deceive the neural network. However, the stop sign is a flat, stationary surface. Researchers wanted to design a surface that was irregular, twisted, and warps with a T-shirt and still could deceive the neural networks.
Researchers wanted to design a shirt while people walk and use that in mathematical problems to model how the shape changes. Lin, along with his colleagues, recorded when a person was walking and used a checkboard pattern to do so. By tracking the corners of the square, they were to observe how the shirt is wrinkling and could model it before printing.
The outcome of this research was a T-shirt that kept the wearer from being spotted. The ultimate goal of this research was deep learning in neural networks and the first step was to analyse the vulnerabilities that could only be fulfilled by this research.
Shweta Tripathi
コメント