Facial recognition problems causing a potential security breach
As you know, we are all a little bit worried about the amount of surveillance that becomes possible when you mix omnipresent cameras with the reliable facial recognition. However, there is a new study out that suggests that some of the best methods are far from failure when it comes to sorting through millions of faces.
At the University of Washington, their real Challenge is an open competition amongst public facial recognition methods that has been running since towards the end of last year.
The prone idea is to observe and see how systems that can outperform humans on sets of thousands of images do when the database size has dramatically increased.
Researchers testing facial recognition
Researchers have started to test with existing labeled image sets of people one of the sets they are testing is consisting of celebrities from a variety of different face angles, and another one tested solely on the individual’s ages.
What the researchers also observed was is with a bit of noise added to the signal in the case of distractors, faces scraped from Creative Commons licensed photos on Flickr.
They ran this test with multiple distractors; something sticks had stayed the same, but they piled up more and more on the hay. What the results had shown was it had a few surprisingly enticing methods, the clear victor for the age-varied set is Google’s FaceNet, while Google and Russia’s N-TechLab are neck and neck in the celebrity database.
Facebook deepface absence
Conspicuously what is absent is Facebook’s DeepFace, which really will be a serious contender. However as participation is completely voluntary and Facebook hasn’t released their system publicly, its performance on MegaFace remains a deep mystery.
Both of the leaders showed a decline over a period as the more distractors implemented, however on that note the efficacy doesn’t fall off quite as fast as the logarithmic scale on the graphs makes it look.
The high accuracy rate by Google in its FaceNet paper does not make it past 10,000 distractors, and by that time there are well over a million, despite a massive lead, this is not accurate enough to serve much of a purpose for the system to be used.
Statistics of accuracy
However for them to get three out of four right with over a million distractors is still really remarkable, but that success rate wouldn’t be beneficial enough in court or as a security product. This seems to be that we still have ways to go before that surveillance state becomes a real life reality.