Makeup to evade the most advanced facial recognition software: according to a study, in 98% of cases the deception was successful
To cheat technology, sometimes you don’t need to be an expert cracker, you don’t even need to know too much about computers, just know your Achilles heel. In the case of facial recognition software, that weak point is in the highly identifiable parts of the face that algorithms use to determine who is who: If these areas of the face are altered in any way, even if the modification is not profound, the machines have serious problems to recognize the subject in real time.
That is the conclusion reached by a recently published study by the Ben-Gurion University of the Negev of Israel, in which the researchers used conventional makeup, which can be bought in any store, on those areas of the face of the children. subjects to fool algorithms. The result was that, when applying the products to key parts of the face, just over 1% of people were recognized by the software.
Israeli researchers explain that made a heat map to define the most identifiable areas of the face, then they used a digital makeup app to project the best way to apply cosmetics and, finally, a makeup artist executed the plan on the faces of 20 participants.
The study highlights that It was tried to make a makeup as natural as possible, that is, like the one that anyone could wear in their day to day, to measure the software’s ability to identify subjects in a realistic situation. And the result was that in just over 98% of the cases the algorithms were not able to recognize them.
The researchers carried out the experiment in a simulated real-world scenario: a corridor equipped with two cameras that transmitted the image live to a state-of-the-art facial detector. They tested the software’s ability to detect faces without makeup in real time –almost 50% correct- and with makeup distributed at random –34% correct- with acceptable results, but when applying cosmetics strategically in areas previously located as highly identifiable, the hit rate fell to just over 1%.
Several studies have previously shown that facial recognition systems can be digitally fooled, for example by creating master faces. And other research had also proven the ability to fool machines in the physical world by placing different elements on the face, such as hats or glasses. But in this second case, the downside is that these items could draw too much attention from security personnel in guarded areas such as airports or government buildings. In the case of makeup, camouflage is very effective with the algorithm without arousing people’s suspicions.