AI is smart enough to recognize your face even if its blurred or pixelated
Artificial Intelligence
Continue reading
Hiding behind pixelated and blurred faces will soon be a thing of the past, thanks to artificial intelligence.

Researchers at the University of Texas at Austin and Cornell Tech has come out with a paper which outlines the working of a software that can reconstruct faces and objects that have been masked using techniques like blurring and pixelation. A blurred number plate or a pixelated human face, anything can be decoded by this image recognition software.

What's even more scary is that the researchers didn't even have to come up with any new image reconstructing algorithms to make it work. Instead, the team used mainstream machine learning methods where the computer is trained with a set of example data rather than programming.

The software was able to beat three proprietary privacy-masking tech, including YouTube's blur tool which YouTube provides to uploaders to select objects and figures they want to blur. The software was able to identify the blurred faces in the videos.

Another method the team unmasked is the method of pixelation, also known as mosaicing. The researchers used standard pixelating techniques found in Photoshop and were able to depixelate the faces in the pictures.

Finally, the team attacked Privacy Preserving Photo Sharing or P3, a tool that is used to encrypt JPEG photos in a way that humans can't see the overall image but other data components are kept intact so computers can still operate on the files. This too was bypassed by the software developed by the researchers.

The software works through trained neural networks that can analyze images from data from four large and well-known image databases. The more faces and objects the neural network 'sees', the better the software gets at recognizing them. Once around 90 per cent accuracy was achieved, the researchers blurred the faces using the three above-mentioned privacy techniques and further trained the AI to recognize blurred and pixelated images.

After that, the researchers used blurred images that the AI hasn't been exposed to yet to see whether it could recognize the images. It was found that the software could recognize 80 to 90 per cent of the blurred images. In case of pixelated faces, the software got success rates of 50 to 75 per cent.

However, its worth noting that the team isn't reconstructing images from scratch. It can't actually reverse the pixelation to recreate the pictures. It can only succeed as long as it knows what it is looking for, things it has seen before.

The larger goal of the researchers with this software is to warn privacy and security enthusiasts that machine learning and AI can be used as a tool for identification and data collection. According to them, there are ways to escape the clutches of AI by using black boxes offering total coverage instead of just distorting the image which leaves certain traces for the AI to pick up and identify.