This post is also available in: heעברית (Hebrew)

Facial recognition technologies are not new. However, what about identifying the actual identity of people under disguise? Researchers from the University of Cambridge are working on an AI-powered facial recognition tech that could soon help identify criminals, political dissidents, protesters, or anyone who conceals their identity by covering their face with masks, hats, scarves or sunglasses.

The technology, dubbed Disguised Face Identification (DFI), uses a deep-learning neural network trained on a dataset of images of people using different props to cover their faces with different backgrounds.

The researchers working on the project are members from the University of Cambridge, National Institute of Technology, and Indian Institute of Science.

How does it work? According to, the technology maps 14 facial points (10 for eyes, one for the nose, and three for lips) on a person’s face and uses the distance and angles between those points to approximate the hidden facial structure. Finally, the system compares the estimated facial structure with learned images to unveil the actual identity of the person in question.

When put to test, the deep-learning algorithm delivered 56% identification accuracy when the face was covered with hats or scarves. With the addition of glasses, the number went down further to 43%, according to a report in Quartz.

Undoubtedly, the AI-based facial-recognition system is still at a nascent stage and will need a number of improvements before being applied practically. The research team understands this need and has released datasets of disguised and undisguised faces, calling on others to test and develop the technology.

Still, the research provides a good insight into the possible applications of a technology that could help identify people just by scanning their masked faces. Law enforcement could be a major beneficiary, but at the same time, it could raise alarms for violating the privacy of a number of people who wear hats and scarves.