Boffins' blur-busting face recognition can ID you with one bad photo
Developers warn that scary people are out there doing this already
Scientists have found a way to accurately identify completely obscured faces using recognition systems trained on only a handful of well-lit photos.
The work by Seong Joon Oh, Rodrigo Benenson, Mario Fritz, and Bernt Schiele of Max Planck Institute in Saarbrücken, Germany, finds faces can be recognised with up to 91.5 per cent accuracy when the system is fed with just 10 clear images of a target's face.
The Faceless Person Recogniser is up to 69.6 per cent accurate when working from just one image.
Accuracy sharply falls when imperfect training images are used. The team introduced black colour into the images dropping performance to 14.7 per cent, still better than random guessing which clocks in at 4.65 per cent.
They warn such an accurate system would likely be already in use.
"It is very probable that undisclosed systems similar to the ones described here already operate online," the team says.
"We believe it is the responsibility of the computer vision community to quantify, and disseminate the privacy implications of the images users share online."
The paper Faceless Person Recognition; Privacy Implications in Social Media [PDF] finds that obfuscation methods including Gaussian blurring are not enough to prevent obscured photos from being used in facial recognition.
Here's the authors' thoughts on the implications of their work:
"From a privacy perspective, the results presented here should raise concern. We show that, when using state of the art techniques, blurring a head has limited effect. We also show that only a handful of tagged heads are enough to enable recognition, even across different events (different day, clothes, poses, point of view).
In the most aggressive scenario considered (all user heads blacked-out, tagged images from a different event), the recognition accuracy of our system is 12× higher than chance level."
The system builds on prior work to fuses multiple convolutional neural network cues from a person's head, body, and "full scene", along with some photo-album data.
Facebook already identifies obscured faces with some 83 per cent accuracy using different and more limited data sets.
Tests were run over the People In Photos album contained 40,000 images and 2000 identities.
"This work is a first step in this direction," the team says. "We conclude by discussing some future challenges and directions on privacy implications of social visual media."
It is the latest step in significant advances in facial recognition over recent years. Facial matching was highly sensitive and failed often when images were of low quality or taken at acute angles such as CCTV. ®
- Black Hat
- Common Vulnerability Scoring System
- Cybersecurity and Infrastructure Security Agency
- Cybersecurity Information Sharing Act
- Data Breach
- Data Protection
- Data Theft
- Digital certificate
- Identity Theft
- Kenna Security
- Palo Alto Networks
- Trusted Platform Module
- Zero trust