Facial recognition already posed serious problems for privacy advocates. Used by everyone from law enforcement to churches, the privacy concerns with facial recognition are very real, and they’re about to get a lot worse.
The ability to identify anyone just by analyzing an image of their face creates a severe imbalance of power from the common citizen to the people in charge. The ability to identify those whose faces are blurred or otherwise obstructed kills that balance entirely. Yet that’s exactly what algorithms like the ‘Faceless Recognition System’ (FRS) are aiming to do.
FRS was a creation by researchers at the Max Planck Institute in Saarbrücken, Germany. The idea was to create a method of identifying individuals through use of imperfect — blurry or otherwise obscured — images. The system trains a neural network on a set of photos containing obscured and unobscured images before using that training to spot similarities from a target’s head and body.
It’s crazy accurate too. The algorithm is able to find an obscured face after seeing an unobscured version of the same face only once at a 69.6 percent accuracy rate. If the machine has 10 images of the person’s face, the accuracy rate climbs to 91.5 percent.
There are, however, limitations. For example, black boxes obscuring a person’s face drop the accuracy rating down to about 14.7 percent, but even that is three times more accurate than humans.
It’s not just one algorithm, either. Facebook has its own facial recognition algorithms that can reportedly identify users with obscured faces at an 83 percent accuracy rate. To do so, it uses cues such as stance and body type. The Faceless Recognition system, however, might be the first fully trainable system that uses a full range of body cues to correctly identify targets.
The researchers recognize the privacy concerns:
From a privacy perspective, the results presented here should raise concern. It is very probable that undisclosed systems similar to the ones described here already operate online. We believe it is the responsibility of the computer vision community to quantify, and disseminate the privacy implications of the images users share online.
In theory, the statement is a good one. The community should police its own creations.
In practice, however, it’s only a matter of time — if it hasn’t happened already — before these algorithms end up in the hands of government, law enforcement and military around the world. At that point, we’re all living in a non-hyperbolic version of 1984.
Siamese neural network.
Siamese neural network is a class of neural network architectures that contain two or more identical subnetworks. identical here means they have the same configuration with the same parameters and weights. Parameter updating is mirrored across both subnetworks.
Siamese NNs are popular among tasks that involve finding similarity or a relationship between two comparable things. Some examples are paraphrase scoring, where the inputs are two sentences and the output is a score of how similar they are; or signature verification, where figure out whether two signatures are from the same person. Generally, in such tasks, two identical subnetworks are used to process the two inputs, and another module will take their outputs and produce the final output. The picture below is from Bromley et al (1993). They proposed a Siamese architecture for the signature verification task.
Siamese architectures are good in these tasks because
- Sharing weights across subnetworks means fewer parameters to train for, which in turn means less data required and less tendency to overfit.
- Each subnetwork essentially produces a representation of its input. (“Signature Feature Vector” in the picture.) If your inputs are of the same kind, like matching two sentences or matching two pictures, it makes sense to use similar model to process similar inputs. This way you have representation vectors with the same semantics, making them easier to compare.
In Question Answering, some recent studies have used Siamese architectures to score relevance between a question and an answer candidate. So one input is a question sentence, the other input is an answer, and the output is how relevant is the answer to the question. Questions and answers don’t look exactly the same, but if the goal is to extract the similarity or a connection between them, a Siamese architecture can work well, too.