Scientists Seong Joon On, Rodrigo Benenson, Mario Fritz and Bernt Schiele from the Max Planck Institute in Saarbrücken, Germany have discovered techniques that may improve the facial recognition technologies used by major social networks such as Facebook. Their research shows that faces can be recognized with 91.5% accuracy when the specialized software is given ten clear images of the person’s face. Improvements are noted when using “protected” photographs of people in which the portraits are partly hidden. Such techniques raise serious privacy concerns.
The research team notes that the facial recognition software that is used on popular social networks such as Facebook use different data sets to provide a joint analysis about the perceived identity of the user. Their findings are published in a paper titled “Faceless Person Recognition; Privacy Implications in Social Media”. The social media may also use meta data extracted from the uploaded images to maximize the accuracy of the face identification process.
The scientists have created an experimental test setup that features a custom made “Faceless Person Recognizer” which proved very accurate when working with obscured faces. The software has been optimized to work with a variety of different patterns of blur that are often employed by people or services to provide anonymity online. The team has discovered that when the program is tested on a few well-lit photographs, then it can correctly identify the person in blurred photos.
The experimental software can detect with a 69.6% accuracy blurred photos when the program has been fed with just a single clear image of the person. When an imperfect photograph is used the accuracy ratio dropped to 14.7% which is still significantly higher than the random guessing technique, which is used by the popular social networks and gives only about 4.65% accuracy.
Such results provide reasonable privacy concerns for all social media users and other platforms and services that use face recognition technologies. The researchers also note that “It is very probable that undisclosed systems similar to the ones described here already operate online.”
The paper also discusses the fact that the popular Gaussian type blurring is not enough to prevent photographs from being identified. The technique that the scientists use shows that when at least several clear photographs of a single user are given to the experimental program, there is a high chance of correct identification of obscured images.
As many more users continue to use social networks and other services that publicly exposes images of themselves along with meta data such containing dates, locations and other content, the privacy risks grow higher.
Reports indicate that Facebook correctly identifies obscured faces with about 83% accuracy using varied data sets.